Hugging Face RL-adapts Qwen for the pi coding agent, doubling SWE-bench Pro pass@1 to 62.7%
victormustar · x · 2026-11-19
Hugging Face's Clement Delangue introduced Qwen3.8-27B-pi-adapted: Qwen's model adapted specifically for the pi coding agent via reinforcement learning and a LoRA adapter.
Measured as pass@1: SWE-bench Pro (Hard 51) jumped from 29.4% to 62.7%, Terminal-Bench 2.1 from 39.5% to 55.8%, with 55–62% fewer tool calls per task.
The release showcases the power of open models — you can adapt them to your own tools and use cases. Open weights and methodology are both published.
Related event: Hugging Face Boosts Qwen to 62.7% on SWE-bench Pro with RL and LoRA(13 posts)→
More from coding & agent
- Free GitHub "AI university" ships a 7-week path from Docker to production RAG and LangGraph agents — mdancho84 · 2026-11-19
- Open-source lemo-animi Skill: one prompt turns Codex into a claymation short-film studio — xiaohu · 2026-11-19
- Vercel Sandbox adds managed npm credentials so coding agents can install private packages safely — cramforce · 2026-11-19
- Authorization Drift benchmark finds frontier LLM agents exceed authorization 43-61% of the time — LuizaJarovsky · 2026-11-19
- ScrapeBench: frontier AI agents ignore scraping bans by default, compliance below 1% without legal prompts — LuizaJarovsky · 2026-11-19
- Developer builds a bootable OS that just runs Claude, on his old MacBook Air, in a day — LukeW · 2026-11-19