Molt: NVIDIA slims down RL code, and the benchmark ends in a draw

On 22 July 2026 NVIDIA's NeMo team posted the paper 'Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning' to arXiv, alongside the release of the code in the official repository under an Apache 2.0 licence. The project presents itself as a 'minimalist' alternative to existing architectures, claiming a codebase of roughly 8,600 lines for the RL path (according to the paper) or 9,200 (according to the README), against some 62,000 lines in verl and 25,000 in slime; OpenRLHF, at around 7,200, stays smaller still. The stated goal is a stack a single researcher — or an AI assistant — can hold in their head, making the algorithmic flow traceable end to end without tens of thousands of lines of distributed glue.

Compactness, however, does not translate automatically into performance supremacy. The headline benchmark, run on 16 H100 GPUs split between training and rollout with a 16,000-token context, shows Molt completing an optimisation step in 119.4 ± 2.3 seconds (461 tokens per GPU per second) against slime's 109.5 ± 10.3 seconds (502 tokens per GPU per second). The authors call the results 'statistically comparable', citing the overlap of the confidence intervals, but the raw figures point to a lower median for the NVIDIA solution. With no independent third-party reproductions, the comparison remains that of a single configuration, measured by the people who signed the framework.

Molt's architecture is built on a native Python interface, using vLLM as the rollout engine and FSDP2 to distribute the policy over NeMo AutoModel. The authors claim the trainer never sees a token the model did not generate, with consistency across tokens, policy versions and model semantics. The paper argues that the same loop stays unchanged from a dense 4-billion-parameter model up to a 700-billion mixture-of-experts policy with expert parallelism 256; for that scale, though, it reports the loop's scalability and not a full comparison against other stacks. Despite these technical credentials, the official documentation states outright that Molt is not a tool for production deployment, pointing to verl or NeMo RL for commercial scenarios that need sustained throughput.

— Olya Watching a hardware vendor bet on clean software to make research easier is a clear-eyed strategic move: if the infrastructure turns into a maze, demand for compute slows down. That a technical draw gets presented as a success reminds us how high the barrier to entry sits today, just to try out a new idea.