Benchmark result

Qwen3.8-Flash-Next-oQ5e-mtp on Apple M5 Max — 64.6 tok/s

Measured with oMLX 0.7.1.dev1 on October 13, 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

3rd fastest of 6 runs of this model on Apple M5 Max · median 64.2 tok/s.

sort

186.1tok/s84.9oMLX · oQ5e · KV tq8
285.9tok/s87.3oMLX · oQ5e · KV tq8
364.6tok/s83.0oMLX · oQ5e · KV tq8this run
463.7tok/s83.6oMLX · oQ5e · KV none
562.4tok/s84.2oMLX · oQ5e · KV none
658.5tok/s80.6oMLX · oQ5e · KV none

Compare the nearest 5 →

Qwen3.8-Flash-Next-oQ5e-mtp on other hardware

Reproduce this run

tooloMLX 0.7.1.dev1

modelQwen3.8-Flash-Next-oQ5e-mtp (oQ5e)

context65,536

kv cachetq8

thinkingon

effortmedium

think limitnone

samplingtemperature 1 · top_p 0.95 · top_k 0 · repetition_penalty 1

clientv0.4.92+102

Set those in oMLX, then:

llm-benchmark benchmark --model "Qwen3.8-Flash-Next-oQ5e-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.