AMD Strix Halo
7 records across 2 models (quants: Gyro-M, Gyro-S, MXFP4).
| hardware | model | quant | backend | tok/s | pp tok/s | W | ttft s | ctx/tg | date | class | flags | source |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| AMD Strix Halo | Qwen3.8-27B | MXFP4 | vLLM | 28.89 | – | – | – | 128 | 2026-10-07 | community | – | source |
| AMD Strix Halo | Qwen3.8-27B | MXFP4 | vLLM | 26.19 | – | – | – | 128 | 2026-10-07 | community | – | source |
| AMD Strix Halo | Qwen3.8-27B | MXFP4 | vLLM | 4.94 | – | – | – | 128 | 2026-10-07 | community | – | source |
| AMD Strix Halo | Qwen3.8-Flash-Next agentionai/llama.cpp Vulkan build, MTP draft where noted | Gyro-S | llama.cpp (Vulkan) | 55 | – | – | – | – | 2026-10-02 | community | – | source |
| AMD Strix Halo | Qwen3.8-Flash-Next agentionai/llama.cpp Vulkan build, MTP draft where noted | Gyro-S | llama.cpp (Vulkan) | 50 | – | – | – | – | 2026-10-02 | community | – | source |
| AMD Strix Halo | Qwen3.8-Flash-Next agentionai/llama.cpp Vulkan build (rotor formats), llama-bench batch 1, no drafting | Gyro-S | llama.cpp (Vulkan) | 31.7 | – | – | – | tg128 | 2026-10-02 | community | – | source |
| AMD Strix Halo | Qwen3.8-Flash-Next agentionai/llama.cpp Vulkan build (rotor formats), llama-bench batch 1, no drafting | Gyro-M | llama.cpp (Vulkan) | 26.4 | – | – | – | tg128 | 2026-10-02 | community | – | source |
All rows cite their source. Provenance: community 40018 sourced 1085