Benchmark result
Qwen3.6-35B-A3B-oQ4e-mtp on Apple M4 Pro — 103.9 tok/s
Measured with oMLX 0.7.1rc1 on October 19, 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
7th fastest of 23 runs of this model on Apple M4 Pro · median 101.0 tok/s.
sort
- 1108.7tok/s77.7oMLX · oQ4e · KV none
- 2105.1tok/s71.1oMLX · oQ4e · KV none
- 3104.6tok/s72.6oMLX · oQ4e · KV none
- 4104.2tok/s75.6oMLX · oQ4e · KV none
- 5104.1tok/s70.3oMLX · oQ4e · KV none
- 6104.0tok/s74.9oMLX · oQ4e · KV none
- 7103.9tok/s78.4oMLX · oQ4e · KV nonethis run
- 8102.5tok/s77.1oMLX · oQ4e · KV none
- 9102.5tok/s70.5oMLX · oQ4e · KV none
- 10102.2tok/s63.7oMLX · oQ4e · KV none
- 11101.3tok/s73.7oMLX · oQ4e · KV none
- 12101.0tok/s65.1oMLX · oQ4e · KV none
- 13100.2tok/s64.6oMLX · oQ4e · KV none
- 1499.8tok/s75.0oMLX · oQ4e · KV none
- 1596.0tok/s73.6oMLX · oQ4e · KV none
- 1694.7tok/s67.3oMLX · oQ4e · KV none
- 1791.5tok/s78.2oMLX · oQ4e · KV none
- 1889.7tok/s68.6oMLX · oQ4e · KV none
- 1986.8tok/s59.5oMLX · oQ4e · KV none
- 2085.8tok/s79.9oMLX · KV none
- 2176.9tok/s70.3oMLX · KV none
- 2276.3tok/s72.8oMLX · oQ4e · KV none
- 2373.8tok/s82.5oMLX · oQ4e · KV none
Reproduce this run
tooloMLX 0.7.1rc1
modelQwen3.6-35B-A3B-oQ4e-mtp (oQ4e)
context131,072
kv cachenone
thinkingon
effortmedium
think limitnone
samplingtemperature 0.7 · top_p 0.8 · top_k 20 · min_p 0 · presence_penalty 1.5 · repetition_penalty 1
clientv0.4.96+102
Set those in oMLX, then:
llm-benchmark benchmark --model "Qwen3.6-35B-A3B-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.