Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model
Article
TL;DR
Qwen3.6-27B runs in 16.8GB, delivers flagship coding at 25 tok/s on a 32GB Mac.
Key Takeaways
- Runs at 25 tok/s on M5 Pro; fits comfortably on 32GB RAM or 24GB VRAM machines
- Noticeable improvement over 3.5-27B; threatens paid Claude plan value for local coders
- Wait 2 weeks before judging — early quants routinely have bugs the community irons out
Discussion
Top comments:
-
[simonw]: Ran it on M5 Pro 128GB; only needs ~20GB, better pelican than Opus 4.7
I ran it on an M5 Pro with 128GB of RAM, but it only needs ~20GB of that. I expect it will run OK on a 32GB machine.
- [syntaxing]: Covers 95% of real coding needs fully local on M4 MBP
- [originalvichy]: Wait weeks for community to find inference backend bugs before judging quality
- [jameson]: Open-source models at fraction of Anthropic pricing erode closed-model moat