Molt: PyTorch-native RL framework matches Megatron performance with cleaner code

Molt is a new open-source training framework for agentic reinforcement learning, designed to be compact and PyTorch-native. Built by Jian Hu and 10 co-authors including Jan Kautz from NVIDIA, Molt addresses the chronic complexity of RL research stacks where every algorithm change requires modifying layers of trainer, distributed backend, and rollout glue. The framework is lean enough for a researcher to hold the entire codebase in memory, and for an AI coding assistant to read and reason about it end to end. This allows algorithm flows to be traced and changed directly. The agent is treated as an ordinary program, and an asynchronous loop trains multimodal and mixture-of-experts (MoE) policies with a guarantee that no token is trained on unless the model itself generated it, ensuring consistency in tokens, policy versions, and semantics.

Performance-wise, Molt does not sacrifice efficiency for simplicity. Under a matched fully asynchronous protocol, its throughput and convergence are statistically comparable to a state-of-the-art Megatron-based stack, a framework typically used for large-scale LLM training. This means researchers get the scalability of a distributed system without the overhead of a complex orchestration layer. Molt is released with recipes and containerized deployments, making it easy to replicate results and build upon. For professionals working with AI agents—such as robotics, game AI, or tool-use systems—Molt offers a practical path to faster iteration on RL training, potentially accelerating breakthroughs in autonomous decision-making.