RL-tuned Qwen3.8-27B for opencode hits Mac via MLX at 156 tok/s on M5 Max

yb2698 · x · 2026-11-19

NetProjects released MLX builds (4/6/8-bit + bf16) of Qwen3.8-27B-opencode-adapted, a Qwen3.8-27B model RL-tuned via harness-hub for the opencode coding agent, now on Hugging Face. Benchmarked with Splash by inco.ai, the 4-bit build (16.3 GB) reaches up to 156 tok/s on an M5 Max, enabling local coding-agent use on Macs.

Related event: Qwen3.8-27B coding model ported to MLX for Mac(2 posts)→

Original post →

More from coding & agent

coding & agent channel →