Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning

NVIDIA NeMo Labs presents Molt, a PyTorch-native framework for agentic reinforcement learning. Its design keeps the codebase compact and makes the agent an ordinary program, while one asynchronous loop supports multimodal and mixture-of-experts policies. The framework aims to preserve consistency across generated tokens, policy versions, and model semantics, including a rule that training never uses tokens the policy did not generate. According to the paper abstract, Molt is statistically comparable to a state-of-the-art Megatron-based stack under a matched, fully asynchronous protocol. Recipes and containers are available as open source.