Qwen3.8-27B-pi LoRA adapter released for the pi coding agent with paired benchmark gains
Original title: harness-hub/Qwen3.8-27B-pi-adapted · Hugging Face
A LoRA adapter trained with AsyncGRPO to serve as the model behind the pi coding agent was released for Qwen3.8-27B, along with full serving and training details. Under pi, Terminal-Bench 2.1 improved from 39.5% to 55.8% and SWE-bench Pro (hard 51) from 29.4% to 62.7%, with 55-64% fewer tool calls per task. The adapter works with vLLM, transformers, llama.cpp GGUF and Llama.app, and the training recipe is available in TRL.
The card provides paired benchmark numbers and full deployment recipes, so readers can judge whether reproducing this RL-trained harness adapter locally is worthwhile.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.