Hugging Face Releases Qwen Adapter for Coding Agents

Hugging Face Releases Qwen Adapter for Coding Agents
Friday, November 20, 2026
  • •Hugging Face released a Qwen3.8-27B LoRA adapter tailored to the open-source coding agent Pi.
  • •Success on 51 challenging tasks rose from 39.2% to 64.7%, while tool calls fell.
  • •The team used its own evaluation, with one attempt per task in the Pi environment.

Hugging Face released Qwen3.8-27B-pi-adapted on the 18th local time, a LoRA adapter that tailors Alibaba’s Qwen3.8-27B model to Pi, an open-source coding agent. The development team said the adapter raised the success rate on 51 challenging SWE-bench Pro tasks from the base model’s 39.2% to 64.7%, while reducing the tool calls needed for tasks.

A LoRA adapter keeps a base model’s structure and weights while training a small set of additional weights to adapt it to a specific task or tool. The team used reinforcement learning (training through feedback and rewards) to train the adapter on repeated coding and terminal tasks. Rewards reflected both whether tasks were solved and how efficiently tools were used.

Both models were evaluated with reasoning mode turned off. Success on the 51 challenging SWE-bench Pro tasks rose from 39.2% to 64.7%, while scores on Terminal-Bench 2.1, which measures terminal task ability, increased from 39.5% to 55.8%. Among tasks both models solved, the median number of tool calls fell from 59 to 23 on SWE-bench Pro and from 18 to 8 on Terminal-Bench 2.1.

The team said the adapter with reasoning mode off achieved a success rate similar to the base model at its highest reasoning setting. In that comparison, the median number of generated tokens per task was about one-fourth as high. To prevent models from finding benchmark answers in advance or manipulating tests to earn rewards, the team also used SafeAgents, which checks all tool calls against rules and blocks requests that could enable cheating.

The adapter is available under the Apache 2.0 open-source license and can run with tools including vLLM and llama.cpp. Hugging Face said it plans to make adapters for other coding agent environments, including OpenCode and Hermes, to use one base model across multiple development tasks. The figures come from the team’s own evaluation, with one attempt per task in the Pi environment; whether the same gains hold in other agent environments or evaluation conditions has yet to be verified.

Subscribe to the AIB Newsletter

The fastest, most reliable AI news, straight to your inbox

Companies related to this article