Qwen3.8-27B opencode-adapted ported to MLX: up to 73 tok/s on M5 Max
yb2698 · x · 2026-11-19
The community ported harness-hub/Qwen3.8-27B-opencode-adapted — a coding-tuned Qwen model adapted for opencode — to MLX with 4/6/8-bit and bf16 weights on Hugging Face. The 4-bit version is just 16.3GB; on an M5 Max it hits 33 tok/s plain, 62 tok/s with MTP, and 73 tok/s with DSpine, making it a strong option for local coding inference on Macs.
More from coding & agent
- Kimi K3.1 Pro beats K3 on 14 of 17 benchmarks; pairing with Opus 5.5 halves spend — Kimi_Moonshot · 2026-11-19
- Apple open-sources .keyx format spec and Python library for Keynote documents — jasonwuishere · 2026-11-19
- 10 open-source GitHub repos that run sales, support, finance and marketing agents for you — victor_explore · 2026-11-19
- Simular box hits OSWorld v2 SOTA at 80.7% success, $1.79 per task — 8x cheaper than August — xwang_lk · 2026-11-19
- Free 4-week course on production AI agents orchestrating GPU media generation draws ~1500 learners — Al_Grigor · 2026-11-19
- From $180 first gig to $20k AI installs: two pricing plays for selling AI to mid-size firms — aryanXmahajan · 2026-11-19