Skip to content

Hugging Face releases RL-tuned LoRA adapters for coding agents

Short answer

Hugging Face released LoRA adapters for Qwen3.8-27B, each trained with reinforcement learning inside one agent harness (pi, OpenCode or Hermes Agent), plus a merged super-adapter. Under pi, Terminal-Bench 2.1 pass@1 rose from 39% to 55% while average tokens per task fell from 2.9K to 1.5K.

What this means for operators

This only matters if your team already runs a coding or ops agent on an open model such as Qwen3.8-27B inside pi, OpenCode or Hermes Agent. In that case the adapters are a ready-made way to cut tokens and tool calls per task, which is where agent cost and latency come from. Before adopting one, rerun it on a sample of your own tasks and compare cost per solved task. The published numbers come from coding benchmarks, and the source does not test sales, support or back-office workflows. The bigger lesson for a 10-200 person company is the recipe: tuning a frozen model to your own harness and task mix can pay off more than switching models.

Hugging Face has published a set of harness adapters for Qwen3.8-27B. They are small LoRA layers trained with reinforcement learning inside the agent harness that will run the model. The team trained one per harness (pi, Hermes Agent and OpenCode) and then merged the three into a single super-adapter. Adapters, GGUF files, the RL tasks and a safeagents plugin are collected on the Hub.

What the adapters do

The base weights stay frozen. The adapter changes how the model behaves in one harness: it reasons less, calls fewer tools and gets more tasks right on the first attempt. The headline example is pi on Terminal-Bench 2.1. The base model scores 39% pass@1 and uses about 2.9K tokens per task. With the adapter it reaches 55% at about 1.5K tokens.

Training setup

Training used TRL's AsyncGRPOTrainer, OpenEnv and Harbor, with rollouts running in Hugging Face Jobs. The harnesses ran as shipped, and the trainer learned from the tokens they actually produced. The task set was filtered from 1,127 Hub candidates, of which 858 passed validation. The published dataset holds 871 tasks from 35 repositories, and 171 of them are "qualified" because the base model solves them only some of the time. The three runs covered 598 steps in total. The safeagents tool checked about 1.3M tool calls and blocked 647 of them.

Super-adapter results

The three adapters were merged at half weight each. With thinking off, mean pass@1 across six harness and benchmark pairs rose from 39.5% to 49.0%. The merged adapter made 53-73% fewer tool calls and generated 44-60% fewer tokens on tasks both models solved. With thinking on, it beat the base model at the same effort on five of six pairs and tied the sixth. Tool calls fell 31-48% on SWE-bench Pro, and on Terminal-Bench they fell 19-25% with token counts about flat.

Caveats from the source

  • Each harness's own adapter was ahead on four of six pairs with thinking off, so a team using one harness should take that specialist.
  • Results cannot be compared across harnesses, because default settings differ, for example in how earlier reasoning turns are kept.
  • The merge weight was chosen on Terminal-Bench 2.1, so those figures are slightly optimistic.
  • Evaluation used one attempt per task, covering all 89 Terminal-Bench 2.1 tasks and the 51 hardest SWE-bench Pro tasks.
  • The source text available to us was cut off partway through the evaluation section.

Source: Hugging Face

INITE data

What we measured ourselves on the tools in this item. You will not find it in the source.

Data as of Nov 20, 2026

The tools in this story

  • OpenCode - price, who uses it, alternatives
Next step

Visibility Analyzer

An AI news item will not tell you how assistants see your own site. The Visibility Analyzer checks that on your live site: which answers cite you, which pages an engine cannot retrieve, and what to fix first. Free to run.

Run a free visibility audit

Free to run. No card.

Fee
Free
Length
One run, minutes

Free tier: two analyses a day, no card required.