Qwen3.8 27B
on M2 Max · oMLX · 8,192 ctx
Summary
User reports Qwen3.8-27B prompt processing at 171.3 t/s on an M2 Max Mac Studio with an Intel Arc 140T iGPU helper, up from 137.8 t/s on the Mac alone at 8k context.
Setup is oMLX with oQ8e 8-bit weights, 22 of 64 layers offloaded to the helper laptop (Core Ultra 9 285H), linked over Wi-Fi 7 MLO. Decode is unchanged by design.
Prefill rises from 129.8 to 180.6 t/s at 16k and from 95.2 to 140.0 t/s at 64k. Single runs, no error bars; output equality is not yet verified and only Qwen3.8 27B has run on real hardware.
Compare this run
Medians of plain runs: one device, one request, no speculative decoding, the whole model in the device's memory. The report above is one setup, so read it beside them and not as one of them.
Qwen3.8 27B at 4-bit by hardware
- NVIDIA RTX 509032 GB80.0 t/s · 7 runs
- 67.0 t/s · one run
- NVIDIA RTX 5070 Ti16 GB54.7 t/s · one run
- 50.0 t/s · one run
- NVIDIA RTX 409024 GB40.0 t/s · 3 runs
- AMD RX 7900 XTX24 GB40.0 t/s · 7 runs
Bars are medians on one scale, 0 to 94.0 tokens/s. Each dot is one run and opens its report. Press ▶ to watch text arrive at a row's speed.