harness-hub releases a Qwen3.8-27B LoRA for the pi agent
Original title: harness-hub/Qwen3.8-27B-pi-adapted — new model trending #29 on Hugging Face
Authors report Terminal-Bench 2.1 at 55.8% versus 39.5% for the base model, with 62% fewer tool calls.
harness-hub published a LoRA adapter for Qwen/Qwen3.8-27B, trained with reinforcement learning to drive the pi coding agent on software-engineering and terminal tasks. The card reports large gains versus the base model under pi on Terminal-Bench 2.1 and SWE-bench Pro hard 51, plus fewer tool calls. Step 160 was chosen from checkpoint scores during training. A full recipe is promised in a later blog post; GGUF weights and vLLM, transformers and llama.cpp serving notes are included.
Unconfirmed: Benchmark gains are the publisher's paired evaluation, not an independent leaderboard result.
- SWE-bench Pro hard 51 moved from 29.4% to 62.7%, the authors say, with 18 newly solved and 1 lost.
- Terminal-Bench 2.1 used 86 tasks; the held-out 68 still showed +16.2 points.
- Runs used one attempt, temperature 0.8, and thinking off.
- A linked GGUF repo is harness-hub/Qwen3.8-27B-pi-GGUF.
- The adapter was not trained or evaluated with thinking enabled.
- Products
- pi
- Organizations
- harness-hub
- AI models
- Qwen3.8-27BQwen3.8-27B-pi-adapted
From the original article
6 min read at huggingface.coA LoRA adapter for [Qwen/Qwen3. 8-27B](https://huggingface. co/Qwen/Qwen3.
By harness-hub · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.