Hugging Face adapts Qwen3.8-27B with RL+LoRA for pi coding agent, doubling SWE-bench Pro scores
ben_burtenshaw · x · 2026-11-18
Hugging Face released harness-hub/Qwen3.8-27B-pi-adapted, adapting Qwen's open model specifically for the pi coding agent via reinforcement learning and a LoRA adapter. It's now #5 trending on the Hub.
Measured pass@1 gains:
- SWE-bench Pro (Hard 51): 29.4% → 62.7%
- Terminal-Bench 2.1: 39.5% → 55.8%
- 55–62% fewer tool calls per task
Open weights and methodology are published. The pitch: the power of open models is adapting them to your own tools and use cases.
More from coding & agent
- Mitchell Hashimoto's Superlogical opens public testing for Rex, a terminal multiplexer built for the AI agent era — dfinke · 2026-11-18
- GitHub Copilot lands in Slack: go from conversation to branch and pull request — mariorod1 · 2026-11-18
- levelsio recounts turning Claude failures into a multiplayer flight sim everyone said was impossible — RealGeneKim · 2026-11-18
- Business owner edits a Google Doc, a daemon sends it to Claude Code to ship WordPress changes — colinmcnamara · 2026-11-18
- Production-grade AI agents are full architectures, not just prompts — goyalshaliniuk · 2026-11-18
- Dev shares a real consumer A2A use case: agents filtering awkward first intros — firdo_dev · 2026-11-18