Don’t use Qwen 3.8 27B on your M5 Max / Ultra
retained text
Not sure people realize this, and why there are so much false information: if you are the lucky few having M5 Max 128G / M5 Ultra you will be much happier with a low-bit DeepSeek 4.1 Flash 1017 than BF16 Qwen 3.8 27B 1. It uses less RAM, with SSD streaming, you can cap the RAM to 50G and still get 30 tok/s 2. It uses less reasoning tokens. 4.1 still uses a lot reasoning tokens, but still 5x fewer than your Qwen 3.8 27B 3. Not quite cloud, but close! You get 4000 tok/s on prefill and 100 tok/s on decode on M5 Ultra, which feels quite usable! 4. No performance degradation on long context: just keep your context going to 200k, 300k, 400k, smartness / performance is the same. What’s your counter argument?!
document fetched 2026-11-12 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xtkswo/dont_use_qwen_38_27b_on_your_m5_max_ultra