huggingface.co web signal

harness-hub Ships Qwen3.8-27B pi-adapted LoRA, Hits 64.7% on SWE-bench Pro Hard and Cuts Tokens 4.5x

Summary

harness-hub released a LoRA adapter for Qwen3.8-27B trained via AsyncGRPO on 750 scale800 + 300 TermGrade tasks across 2× B200 GPUs, built to power the pi coding agent. The step-160 checkpoint pushes SWE-bench Pro Hard 51 from 39.2% to 64.7% (+25.5 points) and Terminal-Bench 2.1 from 39.5% to 55.8%, while generating 4.5× fewer tokens than the base model at equivalent accuracy — all with thinking disabled.