HF CEO: RL-Adapted Qwen3.8-27B Doubles SWE-bench Pro Score to 62.7% for pi Agent
victormustar · x · 2026-11-18
Hugging Face CEO Clement Delangue introduced Qwen3.8-27B-pi-adapted, built by adapting Qwen's open model specifically for the pi coding agent using reinforcement learning (async-GRPO) and a LoRA adapter.
Results (pass@1):
- SWE-bench Pro (Hard 51 subset): 29.4% → 62.7% (+34 pts)
- Terminal-Bench 2.1: 39.5% → 55.8% (+16 pts)
- 55–62% fewer tool calls per task
Open weights are on Hugging Face (harness-hub/Qwen3.8-27B-pi-adapted) with methodology published. The pitch: open models can be adapted to your own tools and use cases.
More from coding & agent
- Neon MCP Has Double the DAU of Its CLI Despite All Efforts Pushing CLI — JensHonack · 2026-11-18
- OpenAI's Agent-Swarm Playbook for Boring Spark Upgrades, Shown at QCon — AdiPolak · 2026-11-18
- Orthogonal Launches Data Marketplace for Agents With 100+ Per-Call Data Providers — ycombinator · 2026-11-18
- Replit CEO on Missions: an Agentic Swarm Harness That Hits Goals Within Budget — amasad · 2026-11-18
- User says he'd pay $20/month extra for an X API that lets agents read posts — willdepue · 2026-11-18
- Perplexity launches Orchestrator API: 45+ models, auto-routing, unified billing — perplexity_ai · 2026-11-18