Benchmark result
Qwen3.8-27B-oQ6e-mtp on Apple M5 Max — 49.4 tok/s
Measured with oMLX 0.7.0.dev4 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
9th fastest of 11 runs of this model on Apple M5 Max · median 50.5 tok/s.
sort
- 161.5tok/s63.5mlx-serve
- 260.5tok/s80.9mlx-serve
- 360.2tok/s85.6mlx-serve
- 459.8tok/s85.3mlx-serve
- 552.2tok/s86.1oMLX · oQ6e · KV none
- 650.5tok/s83.0oMLX · oQ6e · KV none
- 750.4tok/s86.3oMLX · oQ6e · KV none
- 849.5tok/s81.8oMLX · oQ6e · KV none
- 949.4tok/s82.1oMLX · oQ6e · KV nonethis run
- 1049.1tok/s86.6oMLX · oQ6e · KV none
- 1129.2tok/s85.7oMLX
Reproduce this run
tooloMLX 0.7.0.dev4
modelQwen3.8-27B-oQ6e-mtp (oQ6e)
context65,536
kv cachenone
effortmedium
think limitnone
samplingtemperature 1 · top_p 0.95 · top_k 20 · min_p 0 · presence_penalty 1.5 · 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.