Skip to content

How Hugging Face Trained Qwen3.8-27B Adapters for Three Agent Harnesses

Hugging Face BlogFrom the source·· 1 day agoPicksIntroducing harness adapters: adapt your favorite model to your harness to improve performance and cost
Original
Image: Hugging Face Blog
Who
Hugging Face
What
shares a reusable recipe for training RL harness adapters for agentic coding across Pi, OpenCode, and Hermes Agent
Quote
And most of all, share a fully reusable recipe, so you can reproduce this on your own task, harness, model combinations.

What happened

Hugging Face trained separate LoRA adapters for Qwen3.8-27B inside Pi, Hermes Agent, and OpenCode, then combined them into a single adapter. On held-out coding benchmarks, each specialist outperformed the base model in its own harness, with 53–72% fewer tool calls on SWE-bench Pro and 53–70% fewer on Terminal-Bench 2.1 for tasks both models solved. The post shares the training pipeline and artifacts for reproducing the approach.

Why it matters

Training a model inside its target agent harness improved coding benchmark results while reducing tool calls; the released recipe lets teams build adapters for their own model, tasks, and harness.

The story so far

Hugging Face and Harness Hub have released reinforcement-learning-adapted LoRA adapters for Qwen3.8-27B, along with open weights and a reusable training pipeline, to improve coding agent performance across Pi, Hermes Agent, and OpenCode. Harness Hub first released open weights for Qwen3.8-27B-pi-adapted, a version tuned for the pi coding agent. Clément Delangue reported that this model raised pass@1 rates from 29.4% to 62.7% on SWE-bench Pro (Hard 51) and from 39.5% to 55.8% on Terminal-Bench 2.1, while cutting tool calls per task by 55–62%. Hugging Face later detailed how it trained separate LoRA adapters for Qwen3.8-27B inside Pi, Hermes Agent, and OpenCode before merging them into a single adapter. According to Hugging Face, each specialist outperformed the base model in its own harness on held-out coding benchmarks, using 53–72% fewer tool calls on SWE-bench Pro and 53–70% fewer on Terminal-Bench 2.1 for tasks both models solved, and the company shared its training pipeline and artifacts for reproduction.

How the story developed · 2

View full event
  1. 11/20 22:34 (UTC+8)

    How Hugging Face Trained Qwen3.8-27B Adapters for Three Agent Harnesses This report

  2. 11/18 04:13 (UTC+8)
    Qwen3.8-27B-pi-adapted Targets the pi Coding AgentClément Delangue · X

Read next

See the full storyHugging Face and Harness Hub Release RL-Adapted Qwen3.8-27B Coding Agent Models and Training Recipe
This story is in 11/21 AI DailyRead the whole issue