Benchmark result
Qwen3.8-27B-oQ8e-mtp on Apple M5 Max — 39.4 tok/s
Measured with oMLX 0.31.3 on September 9, 2026.
Support llm-bench.io
llm-bench.io is free. If it helped you, I'd appreciate your support to cover the infrastructure costs.
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
27th fastest of 64 runs of this model on Apple M5 Max · median 38.5 tok/s.
sort
- 161.4tok/s70.7oMLX · oQ8e · KV none
- 256.3tok/s84.5oMLX · oQ8e · KV none
- 355.0tok/s74.3oMLX · oQ8e · KV none
- 452.4tok/s84.0oMLX · oQ8e · KV none
- 552.3tok/s84.5oMLX · oQ8e · KV none
- 651.9tok/s85.7oMLX · oQ8e · KV none
- 751.8tok/s88.8oMLX · oQ8e · KV none
- 851.7tok/s89.9oMLX · oQ8e · KV none
- 951.6tok/s79.4oMLX · oQ8e · KV none
- 1051.3tok/s89.9oMLX · oQ8e · KV none
- 1150.8tok/s83.0oMLX · oQ8e · KV none
- 1250.3tok/s87.9oMLX · oQ8e · KV none
- 1349.2tok/s86.9oMLX · oQ8e · KV none
- 1449.0tok/s86.9oMLX · oQ8e · KV none
- 1548.2tok/s85.6oMLX · oQ8e · KV none
- 1647.7tok/s69.9oMLX · oQ8e · KV none
- 1746.6tok/s85.7oMLX · oQ8e · KV none
- 1845.5tok/s83.8oMLX · oQ8e · KV none
- 1944.9tok/s74.5oMLX · oQ8e · KV none
- 2043.4tok/s69.4oMLX · oQ8e · KV none
- 2143.2tok/s86.0oMLX · oQ8e · KV none
- 2242.4tok/s82.0oMLX · oQ8e · KV none
- 2342.3tok/s87.9oMLX · oQ8e · KV none
- 2440.5tok/s82.7oMLX · oQ8e · KV none
- 2539.8tok/s87.4oMLX · oQ8e · KV none
- 2639.7tok/s82.6oMLX · oQ8e · KV none
- 2739.4tok/s78.7oMLX · KV nonethis run
- 2839.4tok/s73.9oMLX · oQ8e · KV none
- 2939.1tok/s75.7oMLX
- 3039.0tok/s81.2oMLX · KV none
- 3138.6tok/s72.0oMLX
- 3238.5tok/s89.6oMLX · oQ8e · KV none
- 3338.5tok/s76.0oMLX
- 3438.2tok/s64.2oMLX · oQ8e · KV none
- 3538.1tok/s85.5oMLX · oQ8e · KV none
- 3637.6tok/s89.0oMLX · oQ8e · KV none
- 3737.6tok/s90.1oMLX · oQ8e · KV none
- 3837.3tok/s88.3oMLX · KV none
- 3936.7tok/s70.0oMLX
- 4035.6tok/s85.6oMLX · oQ8e · KV none
- 4135.4tok/s69.2oMLX
- 4235.0tok/s73.1oMLX
- 4334.4tok/s85.8oMLX
- 4433.9tok/s82.5oMLX
- 4533.7tok/s85.6oMLX
- 4633.7tok/s87.3oMLX · oQ8e · KV none
- 4733.5tok/s85.9oMLX
- 4833.4tok/s87.6oMLX · KV none
- 4933.2tok/s84.0oMLX
- 5032.9tok/s85.7oMLX · oQ8e · KV none
- 5132.7tok/s83.0oMLX
- 5232.4tok/s89.0oMLX
- 5332.3tok/s84.8oMLX
- 5432.2tok/s88.5oMLX
- 5532.1tok/s83.7oMLX · KV none
- 5632.1tok/s81.8oMLX · KV none
- 5731.8tok/s86.2oMLX
- 5830.8tok/s84.7oMLX
- 5930.0tok/s83.9oMLX
- 6029.3tok/s87.7oMLX · KV none
- 6127.5tok/s86.3oMLX · KV tq4
- 6227.3tok/s87.0oMLX
- 6327.1tok/s68.7oMLX
- 6420.4tok/s82.4oMLX
Reproduce this run
tooloMLX 0.31.3
modelQwen3.8-27B-oQ8e-mtp
context262,144
kv cachenone
thinkingon
effortmedium
think limitnone
samplingtemperature 1 · top_p 0.95 · top_k 20 · repetition_penalty 1
clientv0.4.59+97
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.