Benchmark result
Qwen3.8-27B-oQ4e-mtp on Apple M4 Max — 44.5 tok/s
Measured with oMLX 0.32.0 on September 22, 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
50th fastest of 70 runs of this model on Apple M4 Max · median 45.0 tok/s.
sort
- 147.2tok/s83.1oMLX · KV none
- 246.5tok/s80.1oMLX · oQ4e · KV none
- 346.5tok/s83.4oMLX · KV none
- 446.3tok/s84.3oMLX · KV none
- 546.2tok/s82.4oMLX · KV none
- 646.1tok/s84.1oMLX · KV none
- 746.0tok/s82.8oMLX · oQ4e · KV none
- 845.9tok/s75.8oMLX · KV none
- 945.8tok/s86.9oMLX · oQ4e · KV none
- 1045.7tok/s82.8oMLX · KV none
- 1145.7tok/s85.1oMLX · oQ4e · KV none
- 1245.6tok/s86.6oMLX · oQ4e · KV none
- 1345.6tok/s86.6oMLX · oQ4e · KV none
- 1445.6tok/s82.1oMLX · oQ4e · KV none
- 1545.6tok/s84.6oMLX · oQ4e · KV none
- 1645.6tok/s82.6oMLX · KV none
- 1745.6tok/s83.0oMLX · oQ4e · KV none
- 1845.6tok/s85.9oMLX · KV none
- 1945.5tok/s82.7oMLX · KV none
- 2045.5tok/s82.6oMLX · KV none
- 2145.5tok/s87.1oMLX · KV none
- 2245.4tok/s83.2oMLX · oQ4e · KV none
- 2345.4tok/s84.6oMLX · oQ4e · KV none
- 2445.4tok/s85.1oMLX · KV none
- 2545.4tok/s82.9oMLX · oQ4e · KV none
- 2645.3tok/s85.1oMLX · KV none
- 2745.3tok/s84.9oMLX · oQ4e · KV none
- 2845.3tok/s84.7oMLX · KV none
- 2945.2tok/s80.9oMLX · oQ4e · KV none
- 3045.1tok/s83.7oMLX · KV none
- 3145.1tok/s85.6oMLX · oQ4e · KV none
- 3245.0tok/s84.1oMLX · KV none
- 3345.0tok/s86.2oMLX · oQ4e · KV none
- 3445.0tok/s84.0oMLX · KV none
- 3545.0tok/s85.9oMLX · KV none
- 3644.9tok/s87.2oMLX · KV none
- 3744.9tok/s84.5oMLX · KV none
- 3844.9tok/s79.0oMLX · KV none
- 3944.8tok/s78.7oMLX · oQ4e · KV none
- 4044.8tok/s82.3oMLX · oQ4e · KV none
- 4144.7tok/s87.1oMLX · oQ4e · KV none
- 4244.7tok/s75.6oMLX · KV none
- 4344.7tok/s84.5oMLX · oQ4e · KV none
- 4444.7tok/s82.2oMLX · KV none
- 4544.7tok/s80.4oMLX · oQ4e · KV none
- 4644.7tok/s88.6oMLX · KV none
- 4744.6tok/s84.2oMLX · KV none
- 4844.6tok/s77.3oMLX · oQ4e · KV none
- 4944.6tok/s85.5oMLX · oQ4e · KV none
- 5044.5tok/s87.6oMLX · oQ4e · KV nonethis run
- 5144.5tok/s80.6oMLX · oQ4e · KV none
- 5244.4tok/s80.7oMLX · oQ4e · KV none
- 5344.4tok/s84.0oMLX · KV none
- 5444.4tok/s77.8oMLX · KV none
- 5544.4tok/s79.6oMLX · oQ4e · KV none
- 5644.2tok/s84.5oMLX · KV none
- 5744.2tok/s82.3oMLX · oQ4e · KV none
- 5844.2tok/s78.9oMLX · KV none
- 5944.1tok/s86.3oMLX · KV none
- 6044.1tok/s87.4oMLX · KV none
- 6144.0tok/s85.7oMLX · oQ4e · KV none
- 6244.0tok/s83.7oMLX · KV none
- 6344.0tok/s86.3oMLX · KV none
- 6443.9tok/s81.5oMLX · KV none
- 6543.8tok/s84.7oMLX · KV none
- 6643.7tok/s87.2oMLX · KV none
- 6743.7tok/s79.3oMLX · KV none
- 6843.5tok/s79.1oMLX · KV none
- 6943.4tok/s79.5oMLX · KV none
- 7042.5tok/s83.7oMLX · oQ4e · KV none
Reproduce this run
tooloMLX 0.32.0
modelQwen3.8-27B-oQ4e-mtp (oQ4e)
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-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.