Benchmark result
Qwen3.8-27B-4bit on Apple M4 Pro — 34.4 tok/s
Measured with oMLX 0.7.1rc1 on October 20, 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 3 runs of this model on Apple M4 Pro · median 34.4 tok/s.
sort
- 140.0tok/s67.4oMLX · 4bit · KV none
- 234.4tok/s82.5oMLX · 4bit · KV nonethis run
- 312.3tok/s63.5oMLX · 4bit · KV none
Qwen3.8-27B-4bit on other hardware
Reproduce this run
tooloMLX 0.7.1rc1
modelQwen3.8-27B-4bit (4bit)
context131,072
kv cachenone
thinkingon
effortxhigh (model default)
think limitnone
samplingtemperature 1 · top_p 0.95 · top_k 20 · repetition_penalty 1
clientv0.4.96+102
Set those in oMLX, then:
llm-benchmark benchmark --model "Qwen3.8-27B-4bit" --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.