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

- NVIDIA dropped Molt, a PyTorch-native agentic RL framework that’s 8.6K lines of code. That’s roughly seven times smaller than verl, which is convenient for researchers who want to actually understand their own stack instead of drowning in distributed backend glue. It composes Ray, vLLM, and NeMo AutoModel into one asynchronous loop, keeping agents as plain Python while ensuring token-exact trajectories. Throughput matches Megatron stacks, provided you have the hardware: two nodes of eight H100s. If you’re running on consumer GPUs, this isn’t for you; it’s for well-funded labs and enterprise groups with serious compute budgets. The code is Apache 2.0, so you can fork it, but don’t expect it to run on your laptop. This is infrastructure for frontier labs, not a side-hustle for moonboys. Where's my cut?