The Qwen3.8-27B-pi-adapted model is a new PEFT LoRA adapter designed to optimize the Qwen3.8-27B base model for the pi agent toolkit. Released under the Apache-2.0 license, the adapter modifies specific projections including q, k, v, o, gate, up, and down. It is compatible exclusively with the Qwen3.8-27B architecture, which utilizes a dense hybrid structure with 64 layers alternating between Gated DeltaNet linear-attention and Gated Attention blocks.

Performance metrics show significant improvements in coding tasks. On SWE-bench Pro, the adapter solved 15 new tasks while losing 2, and on Terminal-Bench 2.1, it gained 16 and lost 2. Efficiency gains are notable: median turns on SWE-bench Pro dropped from 59 to 23, and on Terminal-Bench 2.1 from 18 to 8. Additionally, the adapter generates approximately 4.5 times fewer tokens compared to the base model when using thinking.

The adapter achieves accuracy levels comparable to the base model with thinking enabled, without the overhead of long reasoning traces. With thinking active, the adapter reaches 66.7% on SWE-bench Pro and 56.6% on Terminal-Bench 2.1. The implementation incorporates the safeagents proxy to prevent reward hacking by blocking prohibited tool calls, such as fetching benchmark answers from the Hub.

The development follows the Hugging Face harness adapters methodology, which supports hosting multiple adapters on a single base model. This allows for the simultaneous deployment of different agents, such as pi and OpenCode, on one server. Technical resources, including the model card, GGUF adapter, and TRL training script, are available for further implementation and deployment.