Benchmark result

Qwen3.8-27B-oQ8e-fp16-mtp on Apple M5 Max — 29.2 tok/s

Measured with oMLX 0.31.3 on August 28, 2026.

What the model built

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

4th fastest of 4 runs of this model on Apple M5 Max · median 35.2 tok/s.

sort

135.4tok/s88.8oMLX · FP16
235.4tok/s85.6oMLX · FP16
335.0tok/s87.1oMLX · FP16
429.2tok/s76.8oMLX · FP16this run

Compare the nearest 5 →

Qwen3.8-27B-oQ8e-fp16-mtp on other hardware

35.5tok/s82.1Apple M2 Ultra · FP16 · KV none
35.3tok/s85.9Apple M2 Ultra · FP16 · KV none
35.3tok/s83.7Apple M2 Ultra · FP16 · KV none
34.9tok/s83.8Apple M2 Ultra · FP16
34.8tok/s64.1Apple M2 Ultra · FP16 · KV none

Reproduce this run

tooloMLX 0.31.3

modelQwen3.8-27B-oQ8e-fp16-mtp (FP16)

context262,144

thinkingon

clientv0.4.44+97

Set those in oMLX, then:

llm-benchmark benchmark --model "Qwen3.8-27B-oQ8e-fp16-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.