What changed
Hugging Face released several 'harness adapters'—small LoRA adapters trained with reinforcement learning (RL) inside unmodified agent harnesses such as pi, opencode, and hermes. These adapters are designed to make local agentic development more efficient and accurate by reducing the number of tool calls while maintaining or improving accuracy on held-out tasks.
What this means in practice
The harness adapters are trained on verifiable tasks derived from open coding sessions shared on the Hugging Face Hub. These tasks are converted into synthetic RL environments, including repair, QA, build, and extend tasks. Each adapter is trained per harness and can be swapped with the base model for local agent use. Safeagents, a proxy between the harness and the model, checks every tool call with rule-based detectors to block reward-hacking attempts, scoring them 0 if blocked. The training runs are streamed live on a dashboard.
Limitations and availability
The adapters are currently available for LFM2.5-24B-A2B and LFM3-1.7B models, with TwIL-8B as an experimental addition. Training is conducted using the GRPO algorithm with TRL's AsyncGRPOTrainer and tree attention for efficient training. The models are trained on HF Jobs with 2×H200 for LFM3-1.7B and 4×H200 for other runs. The total cost of each training run is visible on the live dashboard. The adapters are hosted as LoRA models and can be loaded for inference by users to make the base model more efficient for their use cases. The supplied announcement does not specify the exact availability timeline for all models.
