Benchmark result
Qwen3.8-27B-oQ6e-mtp on Apple M4 Max — 39.6 tok/s
Measured with oMLX 0.32.0 on September 24, 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 26 runs of this model on Apple M4 Max · median 38.7 tok/s.
sort
- 139.8tok/s84.2oMLX · oQ6e · KV none
- 239.6tok/s88.2oMLX · oQ6e · KV nonethis run
- 339.6tok/s85.0oMLX · oQ6e · KV none
- 439.3tok/s82.5oMLX · oQ6e · KV none
- 539.2tok/s84.1oMLX · oQ6e · KV none
- 639.2tok/s85.7oMLX · oQ6e · KV none
- 739.2tok/s82.2oMLX · oQ6e · KV none
- 839.1tok/s67.0oMLX · oQ6e · KV none
- 939.1tok/s87.7oMLX · oQ6e · KV none
- 1039.1tok/s81.5oMLX · oQ6e · KV none
- 1139.0tok/s81.7oMLX · oQ6e · KV none
- 1239.0tok/s84.5oMLX · oQ6e · KV none
- 1338.7tok/s83.0oMLX · oQ6e · KV none
- 1438.6tok/s83.8oMLX · oQ6e · KV none
- 1538.5tok/s83.7oMLX · oQ6e · KV none
- 1638.4tok/s79.9oMLX · oQ6e · KV none
- 1738.3tok/s80.8oMLX · oQ6e · KV none
- 1838.1tok/s88.6oMLX · oQ6e · KV none
- 1938.1tok/s84.4oMLX · oQ6e · KV none
- 2037.9tok/s84.2oMLX · oQ6e · KV none
- 2137.9tok/s80.9oMLX · oQ6e · KV none
- 2237.8tok/s84.7oMLX · oQ6e · KV none
- 2337.8tok/s88.0oMLX · oQ6e · KV none
- 2437.5tok/s84.9oMLX · oQ6e · KV none
- 2537.4tok/s78.2oMLX · oQ6e · KV none
- 2637.3tok/s80.7oMLX · oQ6e · KV none
Qwen3.8-27B-oQ6e-mtp on other hardware
Reproduce this run
tooloMLX 0.32.0
modelQwen3.8-27B-oQ6e-mtp (oQ6e)
context262,144
kv cachenone
effortmedium
think limitnone
samplingtemperature 1 · top_p 0.95 · top_k 20 · repetition_penalty 1
clientv0.4.80+98
Set those in oMLX, then:
llm-benchmark benchmark --model "Qwen3.8-27B-oQ6e-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.