Benchmark result
Qwen3.8-27B-oQ8e-fp16-mtp on Apple M2 Ultra — 35.3 tok/s
Measured with oMLX 0.31.3 on September 3, 2026.
What the model built
Open full screen →The coding scenario asks for a playable game in a single HTML file. This is exactly what the model returned, unedited; the frame only adds a line that tells this page how tall it is.
How this run compares
2nd fastest of 13 runs of this model on Apple M2 Ultra · median 34.6 tok/s.
sort
- 135.5tok/s82.1oMLX · FP16 · KV none
- 235.3tok/s85.9oMLX · FP16 · KV nonethis run
- 335.3tok/s83.7oMLX · FP16 · KV none
- 434.9tok/s83.8oMLX · FP16
- 534.8tok/s64.1oMLX · FP16 · KV none
- 634.7tok/s75.9oMLX · FP16 · KV none
- 734.6tok/s86.6oMLX · FP16 · KV none
- 834.5tok/s83.9oMLX · FP16
- 934.2tok/s83.7oMLX · FP16 · KV none
- 1033.9tok/s70.4oMLX · FP16 · KV none
- 1133.8tok/s81.6oMLX · FP16
- 1231.5tok/s81.6oMLX · FP16 · KV none
- 1329.8tok/s85.6oMLX · FP16
Qwen3.8-27B-oQ8e-fp16-mtp on other hardware
Reproduce this run
tooloMLX 0.31.3
modelQwen3.8-27B-oQ8e-fp16-mtp (FP16)
context262,144
kv cachenone
think limitnone
samplingtemperature 1 · top_p 0.95 · top_k 20 · repetition_penalty 1
clientv0.4.53+97
Set those in oMLX, then:
llm-benchmark benchmark --model "Qwen3.8-27B-oQ8e-fp16-mtp" --tool "oMLX"The client prompts for context length and thinking mode, and for the KV cache dtype on Unsloth Studio. Sampling is left at the model default — the values above are what the tool reported using, not overrides the client sent.