Hugging Face’s harness-hub team has released Qwen3.8-27B-pi-adapted, a LoRA adapter for Qwen3.8-27B trained with reinforcement learning to work inside the pi coding agent, MarkTechPost reports. The point is narrow on purpose: one adapter, for one harness, on one open base model, instead of a general chat fine-tune. The adapter is Apache-2.0. Pi itself is an MIT-licensed agent toolkit with a coding CLI. You can run it in vLLM with LoRA enabled, or in llama.cpp with a GGUF adapter on a Q8_0 base. The LoRA is rank 32 (alpha 64). A merged BF16 checkpoint is about 54 GB. Training used two NVIDIA B200s, and the run was stopped early, at step 160 of 1,000 planned.
The scores are self-reported, all inside the pi harness, with thinking off. On a 51-task hard slice of SWE-bench Pro, pass@1 is said to rise from 39.2% to 64.7%. On Terminal-Bench 2.1, it rises from 39.5% to 55.8%. On tasks both models already solved, median turns dropped sharply, and the adapter is reported to use about 4.5 times fewer tokens than the base model with thinking on, while matching that thinking-on accuracy. The eval sets are small. I have not run the adapter.
Review before you install: read the release notes yourself, back up your config, and try new versions somewhere you can afford to break first.
My take
A small, harness-specific LoRA on an open 27B model is a healthier shape than another all-purpose chat checkpoint. Those gains are still the team’s own numbers, and they only apply inside pi. If you already run Qwen3.8-27B locally, this is worth a careful look. Read the model card, and try it somewhere you can afford to be wrong.
God Bless, and Happy hacking!