Benchmark result
Qwen3.8-27B-oQ4e-mtp on Apple M5 Max — 29.1 tok/s
Measured with oMLX 0.31.4.dev132+g94cdcae13 on October 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
39th fastest of 42 runs of this model on Apple M5 Max · median 43.1 tok/s.
sort
- 163.6tok/s83.3oMLX · oQ4e · KV none
- 262.6tok/s89.2oMLX · oQ4e · KV none
- 360.8tok/s83.9oMLX · oQ4e · KV none
- 460.4tok/s80.5oMLX · oQ4e · KV none
- 556.1tok/s81.7oMLX · oQ4e · KV none
- 655.1tok/s83.5oMLX · KV none
- 754.8tok/s86.6oMLX · KV none
- 854.8tok/s83.0oMLX · KV none
- 954.5tok/s71.4oMLX
- 1053.2tok/s79.9oMLX · KV none
- 1153.0tok/s84.1oMLX · KV none
- 1250.9tok/s79.3oMLX · KV none
- 1349.6tok/s83.9oMLX · KV none
- 1449.3tok/s82.2oMLX
- 1547.1tok/s88.0oMLX · KV none
- 1647.0tok/s69.4oMLX
- 1746.3tok/s85.0oMLX
- 1846.1tok/s84.0oMLX
- 1944.9tok/s87.3oMLX
- 2044.1tok/s85.6oMLX · oQ4e · KV none
- 2143.3tok/s85.4oMLX · KV none
- 2243.0tok/s85.9oMLX
- 2342.2tok/s87.2oMLX
- 2441.2tok/s88.2oMLX
- 2541.2tok/s86.7oMLX · KV none
- 2640.6tok/s84.0oMLX · KV none
- 2740.6tok/s84.0oMLX · KV none
- 2840.4tok/s82.9oMLX · KV none
- 2939.9tok/s83.4oMLX
- 3039.5tok/s85.8oMLX · oQ4e · KV tq8
- 3137.9tok/s83.9oMLX · oQ4e · KV tq4
- 3237.3tok/s85.0oMLX · oQ4e · KV tq6
- 3336.7tok/s86.8oMLX
- 3433.6tok/s84.0oMLX
- 3529.9tok/s81.7oMLX · oQ4e · KV none
- 3629.8tok/s69.4oMLX · oQ4e · KV none
- 3729.6tok/s80.5oMLX · oQ4e · KV none
- 3829.5tok/s85.5oMLX · oQ4e · KV none
- 3929.1tok/s84.5oMLX · oQ4e · KV nonethis run
- 4029.1tok/s85.2oMLX · oQ4e · KV none
- 4128.9tok/s84.5oMLX · oQ4e · KV none
- 4228.3tok/s85.2oMLX
Reproduce this run
tooloMLX 0.31.4.dev132+g94cdcae13
modelQwen3.8-27B-oQ4e-mtp (oQ4e)
context262,144
kv cachenone
thinkingon
effortxhigh
think limitnone
samplingtemperature 1 · top_p 0.95 · top_k 20 · repetition_penalty 1
clientv0.4.85+101
Set those in oMLX, then:
llm-benchmark benchmark --model "Qwen3.8-27B-oQ4e-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.