AnalysisAI ModelsNovember 17, 2026

Hugging Face LoRA adapter lifts Qwen3.8-27B coding scores in pi harness

Read original source →huggingface.co

The 640 MB LoRA (rank 32) on Qwen3.8-27B scores 64.7% on SWE-bench Pro hard-51, up from 39.2%, and 55.8% on Terminal-Bench 2.1, up from 39.5%, with thinking off. It matches the base model's max-effort thinking at ~4.5x fewer tokens; the base model already hits 64.7% on SWE-bench Pro with thinking on.

How this story unfolded

2 days · 3 reports · 7 community posts · 10 of 12 shown

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

More stories today

Open the live feed