NVIDIA Unveils PyTorch-Native Molt Framework for Agentic RL

Reinforcement learning is increasingly used to train AI agents for multi-turn reasoning, tool use and code execution, but researchers often face sprawling software stacks when testing new algorithms or rollout schemes. NVIDIA’s NeMo team designed Molt as a compact, PyTorch-native alternative whose training logic can be traced and modified end to end. The approach favors a single backend and direct interfaces, aiming to accelerate experimentation without giving up distributed performance.

NVIDIA researchers posted Molt’s paper and open-source code on July 22, 2026, releasing the framework under the Apache 2.0 license. Its core reinforcement-learning path contains about 8,600 lines of Python, compared with roughly 62,000 for verl and 25,000 for slime. Built around Ray, vLLM, NeMo AutoModel and FSDP2, Molt targets multi-node H100 clusters and supports multimodal and mixture-of-experts policies. NVIDIA said the same asynchronous training path scaled from a 4-billion-parameter dense model to a 700-billion-parameter MoE model while delivering throughput statistically comparable with a Megatron-based stack.