What matters in AI.

Subscribe

Qwen3.8 coding adapter lifts a test score from 39.2% to 64.7%

The adapter has the Apache-2.0 license, but it is made for one coding agent, pi.

Hugging Face released Qwen3.8-27B-pi-adapted, a LoRA adapter (a small add-on that changes how a model works without a full retrain). It was made with reinforcement learning (learning from reward scores) for the pi coding agent. On SWE-bench Pro hard-51, pass@1 (tasks solved on the first try) goes from 39.2% to 64.7%.

What it is

The adapter is for builders who use the pi coding agent with Qwen3.8. It has the Apache-2.0 license.

The numbers

On Terminal-Bench 2.1, the score goes from 39.5% to 55.8%. On SWE-bench Pro hard-51, it goes from 39.2% to 64.7%. These are the numbers in the release.

What is not known

The release gives no test by other groups. It does not tell how the adapter works with agents other than pi. It gives no cost or time for the training.

Insight: The increase comes from tuning a model to one agent

In October, researchers told that changes to the agent software repair procedure failures. Only changes to model weights repair content failures. This adapter changes the model by training for pi. It looks like a move to make an open model fit one agent, and the increase can possibly be less with other agents. Test it with your own agent first.

Sources

  1. New LoRA Adapter Lifts Qwen3.8 Coding Success From 39% to 65%marktechpost.com
AI MATTER · NEWS · AI MATTER · NEWS ·18 NOV2026

Posted

Learn the terms in this story

Guides for this story

More in Models

All Models news