NVIDIA’s NeMo team has released Molt, a PyTorch-native framework for agentic reinforcement learning built around a simple premise: the code should be compact enough for a researcher to hold in their head. Molt’s RL core comes in at roughly 8.6K lines, measured from each framework’s RL entry point, against about 62K for verl and 25K for slime.
Molt ships under Apache 2.0 with launch scripts and a prebuilt container, and composes Ray for placement, vLLM for rollout and NVIDIA AutoModel with FSDP2 for training. None of the three components is forked, so upstream improvements arrive as a container pin rather than a rebase.
An agent in Molt is an ordinary Python program exporting an AgentRunner, with the reward defined in plain code. The framework supports Gymnasium-style environments and chat agents that talk through a loopback server speaking OpenAI and Anthropic wire protocols, capturing token-exact trajectories so training never touches a token the model did not generate.
For mixture-of-experts policies, Molt replays rollout routing decisions during training so small numerical differences cannot flip expert selection. The shipped recipes assume two nodes of eight H100 GPUs, splitting training and rollout evenly.
NVIDIA positions Molt as research infrastructure for frontier labs and well-funded post-training teams, with the technical report benchmarking it against slime on a Qwen3-30B-A3B test.
