Qwen3.8-27B super-adapter weights are up on Hugging Face with full recipe

yb2698 · x · 2026-11-20

yb2698 encourages readers to train their own harness adapter for their favorite model and harness on their own tasks — the central point of the blog post. Resources:

The mechanism, as previously detailed: the harness adapters' weight updates are nearly orthogonal, so they can be concatenated along the rank dimension into one rank-96 LoRA; with a single harness and thinking off, the specialist adapter remains the better pick.

Related event: yb2698 merges a week of LoRA adapters into one super-adapter, boosting Qwen coding scores by up to 20 points(9 posts)→

Original post →

More from coding & agent

coding & agent channel →