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.