A small LoRA adapter for Alibaba’s Qwen3.8-27B has more than doubled the open model’s score on hard SWE-bench Pro tasks when it runs inside the pi coding agent. The adapter, Qwen3.8-27B-pi-adapted, was published on Hugging Face by the harness-hub organization and was trained with reinforcement learning for that one agent harness.
Quick Take
- SWE-bench Pro (hard 51): 29.4% to 62.7% under pi, with 55% fewer tool calls per task.
- Terminal-Bench 2.1: 39.5% to 55.8% under pi, with 62% fewer tool calls.
- What it is: a rank-32 LoRA adapter on top of Qwen3.8-27B, not a new model.
- Runs on: vLLM, transformers serve and llama.cpp, with a GGUF version for local use.
- Training: AsyncGRPO in TRL, on two NVIDIA B200 GPUs per harness.
In plain English
Coding agents such as pi wrap a language model in a loop that reads files, runs commands and fixes errors. A model that is good at coding in general can still be clumsy inside a specific agent. This adapter is a small add-on file that teaches Qwen3.8-27B how to work inside pi. The result is far more solved tasks with fewer wasted steps, on a model you can run yourself.
What happened
The model card shows a real SWE-bench Pro task where the base model “searches and re-runs snippets until it hits the 100-step limit,” while the adapter finds the function, edits it, runs the tests and fixes the last case in 8 tool calls. The authors say the full recipe will follow in a release blog post so others can build adapters for their own harnesses.
Background
The gap between a model and the harness around it has become a theme in coding AI, as our report on harness versus model benchmarks showed. Qwen3.8-27B is already popular for local work, from offline game builds to agent experiments.
How it works
- Method: reinforcement learning with AsyncGRPO, an experimental trainer in TRL, applied as a LoRA with rank 32 and alpha 64.
- Tasks: 750 scale800 tasks interleaved with 300 TermGrade tasks, run in sandboxes with tool calls screened for safety.
- Reward: task correctness from a verifier, plus a small bonus for using fewer tools.
- Checkpoint: step 160 of a run that was stopped at step 191 of 1,000 planned.
- Setup: thinking stays off, as in training. pi must use the OpenAI completions API, because the adapter was not trained on the Responses API.
Why it matters
For developers, a free adapter brings a local 27B model much closer to hosted coding agents on hard tasks.
For companies, harness-specific adapters are a cheap way to tune open models for internal agents without full fine-tuning.
For agent builders, the published recipe could make “one adapter per harness” a standard pattern.
What’s still open
- Selection effects: the checkpoint was picked partly on SWE-bench Pro scores. On the 68 Terminal-Bench tasks not used for selection, the gain was still 16.2 points.
- The GGUF version has not been re-evaluated, and lower-bit quantizations may drift further.
- Whether the gains transfer to other agent harnesses.
Sources
- harness-hub, “Qwen3.8-27B-pi-adapted” model card on Hugging Face
- harness-hub, “Qwen3.8-27B-pi-adapted-GGUF” model card on Hugging Face
i10x Perspective
The lesson is less about Qwen than about fit: a well-matched adapter can matter as much as a bigger model. That is good news for anyone running open models on their own hardware. More in our Coding section.



