Benchmark result
Qwen3.8-27B-oQ8e-mtp on Apple M5 Max — 42.4 tok/s
Measured with oMLX 0.31.4.dev132+g94cdcae13 on October 10, 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
3rd fastest of 37 runs of this model on Apple M5 Max · median 33.7 tok/s.
sort
- 147.7tok/s69.9oMLX · oQ8e · KV none
- 243.2tok/s86.0oMLX · oQ8e · KV none
- 342.4tok/s82.0oMLX · oQ8e · KV nonethis run
- 442.3tok/s87.9oMLX · oQ8e · KV none
- 539.7tok/s82.6oMLX · oQ8e · KV none
- 639.4tok/s78.7oMLX · KV none
- 739.1tok/s75.7oMLX
- 839.0tok/s81.2oMLX · KV none
- 938.6tok/s72.0oMLX
- 1038.5tok/s76.0oMLX
- 1137.3tok/s88.3oMLX · KV none
- 1236.7tok/s70.0oMLX
- 1335.6tok/s85.6oMLX · oQ8e · KV none
- 1435.4tok/s69.2oMLX
- 1535.0tok/s73.1oMLX
- 1634.4tok/s85.8oMLX
- 1733.9tok/s82.5oMLX
- 1833.7tok/s85.6oMLX
- 1933.7tok/s87.3oMLX · oQ8e · KV none
- 2033.5tok/s85.9oMLX
- 2133.4tok/s87.6oMLX · KV none
- 2233.2tok/s84.0oMLX
- 2332.9tok/s85.7oMLX · oQ8e · KV none
- 2432.7tok/s83.0oMLX
- 2532.4tok/s89.0oMLX
- 2632.3tok/s84.8oMLX
- 2732.2tok/s88.5oMLX
- 2832.1tok/s83.7oMLX · KV none
- 2932.1tok/s81.8oMLX · KV none
- 3031.8tok/s86.2oMLX
- 3130.8tok/s84.7oMLX
- 3230.0tok/s83.9oMLX
- 3329.3tok/s87.7oMLX · KV none
- 3427.5tok/s86.3oMLX · KV tq4
- 3527.3tok/s87.0oMLX
- 3627.1tok/s68.7oMLX
- 3720.4tok/s82.4oMLX
Reproduce this run
tooloMLX 0.31.4.dev132+g94cdcae13
modelQwen3.8-27B-oQ8e-mtp (oQ8e)
context65,536
kv cachenone
effortxhigh (model default)
think limitnone
samplingtemperature 1 · top_p 0.95 · top_k 0 · repetition_penalty 1
clientv0.4.89+102
Set those in oMLX, then:
llm-benchmark benchmark --model "Qwen3.8-27B-oQ8e-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.