New LoRA Adapter Lifts Qwen3.8 Coding Success From 39% to 65%
benchmarksMarkTechPostNov 18, 2026
Hugging Face released Qwen3.8-27B-pi-adapted, an Apache-2.0 LoRA adapter trained with reinforcement learning for the pi coding agent. It lifts SWE-bench Pro hard-51 pass@1 from 39.2% to 64.7% and Terminal-Bench 2.1 from 39.5% to 55.8%. The post New LoRA Ada...
Impact: LowConfidence: LowNovelty: Incremental
benchmarksmarktechpostagentscoding