What happened
Clément Delangue introduced Qwen3.8-27B-pi-adapted, a Qwen model adapted for the pi coding agent with reinforcement learning and a LoRA adapter. The post reports pass@1 increases from 29.4% to 62.7% on SWE-bench Pro (Hard 51) and from 39.5% to 55.8% on Terminal-Bench 2.1, with 55–62% fewer tool calls per task. Open weights and the methodology are linked.
Why it matters
The adapted model improves pass@1 on SWE-bench Pro (Hard 51) and Terminal-Bench 2.1 while using 55–62% fewer tool calls per task, suggesting coding agents can benefit from models tuned to their specific tools.
Introducing Qwen3.8-27B-pi-adapted 🤗
We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter.
The results, measured as pass@1:
🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7%
💻 Terminal-Bench 2.1: 39.5% → 55.8%
⚡ 55–62% fewer tool calls per task
This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛
Open weights: huggingface.co/harness-hub/Qw…
Methodology: huggingface.co/blog/harness-h…
The story so far
Harness Hub has released Qwen3.8-27B-pi-adapted, an iteration of the Qwen model optimized for the pi coding agent using reinforcement learning and a LoRA adapter. The model's weights and methodology are available publicly. Clément Delangue reported that this adaptation improves performance across coding benchmarks. On SWE-bench Pro (Hard 51), pass@1 rates increased from 29.4% to 62.7%, while scores on Terminal-Bench 2.1 rose from 39.5% to 55.8%. Additionally, the model reportedly requires 55–62% fewer tool calls per task.
