Skip to content
ViqusViqus
Navigate
Company
Blog
About Us
Contact
System Status
Enter Viqus Hub

Hugging Face Trains Per-Harness LoRA Adapters That Lift Qwen3.8-27B Agentic Coding

LoRA Adapters Reinforcement Learning GRPO Agentic Coding Terminal-Bench 2.1 Harness Adapters Qwen3.8-27B
November 20, 2026

This summary and analysis were generated by AI from the original article at Hugging Face Blog and may contain errors (how Viqus works). Read the source for full details.

Viqus Verdict Logo Viqus Verdict Logo 4
Practical Recipe, Narrow Scope
Media Hype 3/10
Real Impact 4/10

Article Summary

Hugging Face introduces harness adapters: LoRA adapters trained with reinforcement learning inside a specific agent harness (Pi, OpenCode, Hermes Agent) on a custom task distribution. The authors report that for Qwen3.8-27B, the Terminal-Bench 2.1 score under the Pi harness rises from 39% to 55% pass@1, while average token usage falls from 2.9K to 1.5K per task. They attribute the gains mainly to the model reasoning less, calling fewer tools, and succeeding more often on the first attempt. Training uses TRL's AsyncGRPOTrainer, OpenEnv and Harbor, with the harness running unmodified and the trainer learning from tokens recorded by a proxy. The task set combines SWE-smith, SWE-Gym and Tasksmith-generated tasks, filtered by validation checks and a difficulty screen that keeps tasks the base model solves only sometimes. The team also merged the three adapters into a single super adapter that works across all three harnesses. Artifacts and the recipe are published in a Hugging Face collection. The results are reported by the vendor on its own benchmarks, and the work is an engineering recipe rather than a new base model.

Key Points

  • RL-trained LoRA adapters per agent harness raised Qwen3.8-27B's Terminal-Bench 2.1 pass@1 under Pi from 39% to 55% while reducing average tokens per task.
  • The pipeline filters tasks by validation and by difficulty, keeping only tasks the base model solves sometimes, since tasks always solved or always failed give no GRPO signal.
  • The three harness LoRAs were merged into one super adapter that runs unchanged under Pi, Hermes Agent and OpenCode, and all artifacts are published on the Hub.

Why It Matters

This is a practical recipe for fine-tuning an open model to a specific agent harness and task mix, which is useful for teams building coding agents. The gains are reported on the authors' own benchmarks with a single 27B model, and the results are not independently replicated. It is an incremental engineering advance that reinforces the trend toward harness-aware RL rather than a shift in the field's direction.


You might also be interested in