# Qwen3.8-27B pi-adapted — a LoRA that tunes Qwen3.8-27B for the Pi agent

> An RL-trained LoRA that makes Qwen3.8-27B reach its max-thinking accuracy inside Pi, with thinking switched off.

- Org: harness-hub
- Date: 2026-11-16
- Category: model
- Source: https://huggingface.co/harness-hub/Qwen3.8-27B-pi-adapted
- Canonical page: https://ai-tldr.dev/releases/harness-hub-qwen3-8-27b-pi-adapted/

## In short

Qwen3.8-27B pi-adapted is an Apache-2.0 LoRA adapter trained with RL to drive the Pi coding agent. With thinking off it lifts SWE-bench Pro (hard 51) from 39.2% to 64.7% and Terminal-Bench 2.1 from 39.5% to 55.8%.

## What it is

Qwen3.8-27B pi-adapted is a LoRA adapter from the harness-hub team that turns the open Qwen3.8-27B model into a better engine for the Pi coding agent on software-engineering and terminal tasks. It ships under Apache-2.0 with a GGUF version for local runs, and it sits in the top 10 of Hugging Face's trending models.

## How it works

The adapter was trained with reinforcement learning (async GRPO in TRL) on agent tasks run inside the Pi harness, using Harbor and OpenEnv environments. Training teaches the model to work without its thinking mode: it writes short notes between tool calls and needs fewer than half the turns of the base model. The card reports that turning thinking back on adds no accuracy.

## Why it matters

With thinking off, the harness-hub adapter matches the base model running at its highest reasoning effort on SWE-bench Pro (64.7% vs 64.7%) while generating about 4.5x fewer tokens per task. For people running a local coding agent on a 27B model, that means the same results with far less waiting and compute. All numbers are under the Pi harness only and are not comparable with the official Qwen3.8-27B card.

## Who it's for

developers running local coding agents

## Try it

npm install -g --ignore-scripts @earendil-works/pi-coding-agent, then serve harness-hub/Qwen3.8-27B-pi-adapted on Qwen/Qwen3.8-27B with vLLM

## Links

- [Model Card](https://huggingface.co/harness-hub/Qwen3.8-27B-pi-adapted)
- [GGUF Weights](https://huggingface.co/harness-hub/Qwen3.8-27B-pi-GGUF)
- [Base Model](https://huggingface.co/Qwen/Qwen3.8-27B)
- [Pi Coding Agent](https://github.com/earendil-works/pi)
- [HF Trending](https://huggingface.co/models?sort=trending)
