Researchers at Hugging Face have introduced Molt, a PyTorch-native training framework designed to streamline agentic reinforcement learning (RL) research. The framework aims to reduce the overhead of constant algorithm modifications, new estimators, and pipeline changes that typically require adjustments across multiple layers in mainstream frameworks.
Molt treats the agent as an ordinary program and uses a single asynchronous loop to train multimodal and mixture-of-experts policies. It ensures that no token is trained on unless it was generated by the policy, maintaining consistency in tokens, policy versions, and model semantics.
Despite its lean design, Molt achieves performance statistically comparable to a state-of-the-art Megatron-based stack under a matched, fully asynchronous protocol. The framework is open source and includes recipes and containers available on GitHub.