HF adapts Qwen for pi coding agent with RL+LoRA: SWE-bench Pro Hard jumps 29.4%→62.7%
art_zucker · x · 2026-11-18
Hugging Face CEO Clement Delangue introduced Qwen3.8-27B-pi-adapted: Qwen's model fine-tuned specifically for the pi coding agent using reinforcement learning and a LoRA adapter, with open weights.
Measured pass@1 results:
- SWE-bench Pro (Hard 51): 29.4% → 62.7%
- Terminal-Bench 2.1: 39.5% → 55.8%
- 55–62% fewer tool calls per task
The point: open models can be adapted to YOUR tools and workflows. Commenters speculate DeepSeek applying the same harness could reach frontier-level performance.
More from coding & agent
- 27 vibe-coded Finder clones appeared in a month, most already dead — KaKi_87 · 2026-11-18
- Forrest PKnight Ranks Programming Languages by How Well AI Writes Them — Dan_Jeffries1 · 2026-11-18
- Claude Plus Unreal Engine 5.9 MCP Builds a Game Prototype in One Day — TomLikesRobots · 2026-11-18
- Open-sourced guardrail-middleware puts deterministic policy checkpoints between agents and their tools — MidnightHeavy3064 · 2026-11-18
- The most indispensable agent is the boring one that renames your invoices daily — Winterocean812 · 2026-11-18
- ZeroIDE-MCP lets browser-based AI agents remotely control multiple Windows PCs over MCP — Danywalery · 2026-11-18