DeepSeek-V4-Flash
54 records across 10 hardware.
| hardware | model | quant | backend | tok/s | pp tok/s | W | ttft s | ctx/tg | date | class | flags | source |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 30780 | – | – | – | 16384 | 2026-10-15 | sourced | – | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 30309 | – | – | – | 16384 | 2026-10-15 | sourced | – | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 29873 | – | – | – | 16384 | 2026-10-15 | sourced | – | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 28939 | – | – | – | 16384 | 2026-10-15 | sourced | – | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 6131 | – | – | – | 16384 | 2026-10-15 | sourced | – | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 5249 | – | – | – | 16384 | 2026-10-15 | sourced | contradiction | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 4906 | – | – | – | 16384 | 2026-10-15 | sourced | contradiction | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 4473 | – | – | – | 16384 | 2026-10-15 | sourced | contradiction | source |
| B300 | DeepSeek-V4-Flash MoE | FP4 | vLLM | 1216 | – | – | – | – | 2026-07 | community | – | source |
| B300 | DeepSeek-V4-Flash MoE | FP4 | vLLM | 773 | – | – | – | – | 2026-07 | community | – | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 485.9 | – | – | – | 16384 | 2026-10-15 | sourced | – | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 455.9 | – | – | – | 16384 | 2026-10-15 | sourced | – | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 422.6 | – | – | – | 16384 | 2026-10-15 | sourced | contradiction | source |
| 4x H200 | DeepSeek-V4-Flash | MXFP4 | vLLM | 421.0 | – | – | – | 16384 | 2026-10-15 | sourced | contradiction | source |
| B300 | DeepSeek-V4-Flash MoE | FP4 | vLLM | 344 | – | – | – | – | 2026-07 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 46.1 | 1204.0 | – | – | 1024 | 2026-10-07 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 45.2 | 437.2 | – | – | 1024 | 2026-10-07 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 44.8 | 590.5 | – | – | 1024 | 2026-10-08 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 41.1 | 1474.0 | – | – | 4096 | 2026-10-07 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 40.8 | 1258.0 | – | – | 4096 | 2026-10-08 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 39.5 | 1416.0 | – | – | 8192 | 2026-10-08 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 39.3 | 1362.0 | – | – | 16384 | 2026-10-08 | community | – | source |
| M5 Max | DeepSeek-V4-Flash 284B-A13B | Q2_K | MLX | 39 | – | – | 0.4 | – | 2026-08 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 39.0 | 924.4 | – | – | 4096 | 2026-10-07 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 37.9 | 1278.0 | – | – | 32768 | 2026-10-08 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 36.4 | 1154.0 | – | – | 65536 | 2026-10-08 | community | – | source |
| RTX 6000 Pro (96GB) | DeepSeek-V4-Flash | Q2 | llama.cpp (CUDA) | 35 | – | – | – | – | 2026-05-26 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 34.0 | 950.8 | – | – | 131072 | 2026-10-08 | community | – | source |
| M5 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 33.7 | 808.9 | – | – | 199680 | 2026-10-08 | community | – | source |
| 4x RTX A6000 | DeepSeek-V4-Flash | Q8_K_XL | llama.cpp | 33 | – | – | – | 1048576 | 2026-09 | community | – | source |
| M3 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 28.9 | 525.2 | – | – | 4096 | 2026-10-08 | community | – | source |
| M3 Ultra | DeepSeek-V4-Flash | 4bit | oMLX | 28.0 | 502.2 | – | – | 16384 | 2026-10-08 | community | – | source |
| M3 Max | DeepSeek-V4-Flash | DS4 imatrix | llama.cpp (Metal) | 27.01 | 45.08 | – | – | 32768 | 2026-07-31 | community | – | source |
| M3 Max | DeepSeek-V4-Flash 284B-A13B | Q2_K | MLX | 27 | – | – | 0.6 | – | 2026-08 | community | – | source |
| 3x RTX PRO 4000 Blackwell | DeepSeek-V4-Flash | unspecified | CUDA | 23.1 | 1128 | – | – | – | 2026-10 | community | – | source |
| 3x RTX PRO 4000 Blackwell | DeepSeek-V4-Flash | unspecified | CUDA | 23.0 | 625 | – | – | – | 2026-10 | community | – | source |
| M5 Max | DeepSeek-V4-Flash | 2bit DQ | oMLX | 13.6 | 109.1 | – | – | 1024 | 2026-10-11 | community | – | source |
| M5 Max | DeepSeek-V4-Flash | 2bit DQ | oMLX | 13.6 | 109.1 | – | – | 1024 | 2026-10-11 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 13.57 | – | – | – | 65535 | 2026-10-05 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 13.20 | – | – | – | 262144 | 2026-10-05 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 13.12 | – | – | – | 262144 | 2026-10-05 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 13.03 | – | – | – | 405000 | 2026-10-05 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 12.7 | – | – | – | 65535 | 2026-10-05 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 12.57 | – | – | – | 405000 | 2026-10-05 | community | – | source |
| 4x RTX A6000 | DeepSeek-V4-Flash | Q8_K_XL | llama.cpp | 11.7 | – | – | – | 1048576 | 2026-09 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 10.96 | – | – | – | 1048576 | 2026-10-05 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 9.71 | – | – | – | 1048576 | 2026-10-05 | community | – | source |
| M5 Max | DeepSeek-V4-Flash | 2bit DQ | oMLX | 9.1 | 180.0 | – | – | 4096 | 2026-10-11 | community | – | source |
| M5 Max | DeepSeek-V4-Flash | 2bit DQ | oMLX | 9.1 | 180.0 | – | – | 4096 | 2026-10-11 | community | – | source |
| M3 Ultra | DeepSeek-V4-Flash | 8bit | oMLX | 8.6 | 338.3 | – | – | 1024 | 2026-10-13 | community | – | source |
| M3 Ultra | DeepSeek-V4-Flash | 8bit | oMLX | 8.2 | 339.4 | – | – | 4096 | 2026-10-13 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 7.53 | – | – | – | 1048576 | 2026-10-05 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 7.08 | – | – | – | 405000 | 2026-10-05 | community | – | source |
| 2x RTX 3090 | DeepSeek-V4-Flash 284B-A13B | UD-Q8_K_XL | llama.cpp (CUDA) | 6.88 | – | – | – | 65535 | 2026-10-05 | community | – | source |