New LoRA Adapter Lifts Qwen3.8 Coding Success From 39% to 65%

What happened
Hugging Face released a new LoRA adapter called Qwen3.8-27B-pi-adapted for the Qwen3.8-27B model. This adapter was trained using reinforcement learning to work specifically with the pi coding agent. The release aims to improve coding task success rates when running the agent locally. The adapter is available under the Apache-2.0 license and is designed for efficient use with the pi harness.
Key details
- The adapter raises coding success on SWE-bench Pro (hard 51) from 39.2% to 64.7% with thinking off.
- Median turns to solve tasks on SWE-bench Pro dropped from 59 to 23.
- The adapter uses LoRA rank 32 (alpha 64) and was trained on 2x NVIDIA B200 GPUs.
- Merged BF16 checkpoint size is about 54 GB.
- The adapter only works with Qwen3.8-27B and targets the pi agent harness.
Why it matters
Developers using the pi coding agent with Qwen3.8-27B can now achieve higher success rates and complete tasks more efficiently, using fewer tokens and shorter sessions. This reduces computational costs and makes it easier to run advanced coding agents locally. The change benefits those seeking to optimize local AI coding workflows without needing to fine-tune large base models.
