NVIDIA AI Releases Molt: A PyTorch-Native Agentic Reinforcement Learning Framework

NVIDIA’s NeMo team released Molt, an Apache 2.0 agentic reinforcement learning framework.

Molt stays compact and research-friendly with about 8.6K lines of RL code, composing Ray, vLLM, and NVIDIA AutoModel instead of forking them.

It supports plain Python agents, token-exact loopback handling, and recipes for tool use, code execution, vision-language tasks, and on-policy distillation.

The framework ships with launch scripts and a container, targeting a 2-node, 16-H100 setup split between training and rollout.

Its design aims to reduce correctness risks in token identity, policy-version handling, and MoE routing for large-scale post-training.