AnalysisAI ModelsNovember 17, 2026

Hugging Face LoRA lifts Qwen3.8-27B coding success from 39% to 65%

Read original source →huggingface.co

The harness-hub adapter, a rank-32 LoRA on Qwen3.8-27B trained with RL inside the pi coding agent, scores 64.7% on SWE-bench Pro (hard 51) versus 39.2% for the base model, and 55.8% on Terminal-Bench 2.1 versus 39.5%. With thinking off it matches the base model's highest reasoning effort using about 4.5x fewer tokens.

How this story unfolded

2 days · 3 reports · 4 community posts · 7 of 9 shown

  1. Nov 17
  2. Nov 18
  3. Nov 19

More stories today

Open the live feed