AI NewsProductivityAnnouncement
Hugging Face LoRA adapters beat Qwen3.8-27B inside one coding harness
Hugging Face released small LoRA adapters that beat the open Qwen3.8-27B base model when tied to one coding harness.

Image: Hugging Face
Why it mattersA team that runs one harness gets a faster and cheaper model than the base, from a small open file.
A small fine-tune can beat a bigger model, when both run inside the same coding harness. Hugging Face released a set of LoRA adapters today. Each one takes the same open Qwen3.8-27B model. Each one trains it with reinforcement learning inside one agent harness, on that harness's own tasks. The lead author is Younes Belkada, with nine co-authors on the TRL and open-models teams.
The base is Qwen3.8-27B. The three harnesses are Pi 1.5.2, OpenCode 2.2.7 and Hermes Agent 0.22.6. Each one runs an agent through the same environments. Each one holds state in a different way. So the model best suited to one is not the model best suited to another.
What the specialists do
On Terminal-Bench 2.1, the base Qwen3.8-27B passes 40.0 percent of tasks with the Pi harness. It uses 2.9K tokens on average, Hugging Face says. With the Pi adapter the same base passes 56.5 percent. It uses 1.5K tokens per task.
On SWE-bench Pro Hard, the Pi adapter lifts the base from 40.0 to 64.0 percent. The Hermes adapter lifts the base from 29.4 to 43.1 percent. The OpenCode adapter lifts the base from 35.3 to 51.0 percent.
A super-adapter merges all three along the rank dimension. It is more balanced across harnesses. It also keeps part of the cost saving. Take the tasks the base and the super-adapter both solve, with thinking off. The super-adapter makes 53 to 73 percent fewer tool calls. It also writes 44 to 60 percent fewer tokens per task.
The method and the cost
Each adapter was trained on two NVIDIA B200 GPUs. One ran the vLLM server, one ran the training. Across the three runs Hugging Face reports 598 cumulative steps, 78,114 samples and 167 million trained tokens. The reward is the task verifier plus a 0.1 tool-efficiency bonus.
The dataset is the open harness-hub/agentic-coding-envs set. It holds 871 tasks from 35 repositories. The sources are SWE-smith, SWE-Gym's TaskTrove and Tasksmith. A safety plugin checked 1.3 million tool calls during training. It blocked 647 of them.
Every adapter, the super-adapter, the dataset, the training recipe and the rollout trajectories are on Hugging Face. The Pi adapter is also released as a GGUF file. A local user can load it beside the base model in one command.
The adapter is tied to one harness and one task mix. A team that changes harness loses the gain. The win is practical for a team that has settled on a harness in production. The generalist model has to budget for every tool it may meet, which costs tokens. The specialist knows its tools in advance and plans fewer calls. Hugging Face names two cases that fit. A code reviewer that always sees the same kind of task. A build agent that runs the same tools every time.
Source
Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost, Hugging Face.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

