Benchmark result
Qwen3.8-Flash-Next-oQ4e-mix6-mtp on Apple M4 Max — 62.3 tok/s
Measured with oMLX 0.7.1.dev1 on October 13, 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
10th fastest of 10 runs of this model on Apple M4 Max · median 73.5 tok/s.
sort
- 177.8tok/s83.1oMLX · oQ4e · KV tq8
- 277.2tok/s82.3oMLX · oQ4e · KV tq8
- 376.1tok/s82.0oMLX · oQ4e · KV tq8
- 474.7tok/s79.6oMLX · oQ4e · KV tq8
- 573.7tok/s80.7oMLX · oQ4e · KV tq8
- 673.2tok/s81.8oMLX · oQ4e · KV tq8
- 772.5tok/s80.5oMLX · oQ4e · KV tq8
- 871.6tok/s85.8oMLX · oQ4e · KV tq8
- 969.6tok/s80.7oMLX · oQ4e · KV tq8
- 1062.3tok/s82.8oMLX · oQ4e · KV tq8this run
Reproduce this run
tooloMLX 0.7.1.dev1
modelQwen3.8-Flash-Next-oQ4e-mix6-mtp (oQ4e)
context65,536
kv cachetq8
effortmedium
think limitnone
samplingtemperature 1 · top_p 0.95 · top_k 20 · repetition_penalty 1
clientv0.4.92+102
Set those in oMLX, then:
llm-benchmark benchmark --model "Qwen3.8-Flash-Next-oQ4e-mix6-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.