NVIDIA’s NeMo team released Molt, a PyTorch-native agentic reinforcement learning framework designed to reduce the cost of algorithm iteration. Its codebase is roughly 8.6K lines of RL code, compared to about 62K lines for verl, 25K for slime, and 7.2K for OpenRLHF. Molt ships under Apache 2.0 with launch codes, Slurm scripts, and a prebuilt container, but the paper positions it as research infrastructure, not a production training service. The shipped recipes assume 2 nodes of 8 H100 GPUs, split 8 for training and 8 for rollout, targeting frontier labs, AI startups, enterprise research groups, and academic labs with multi-node H100/H200 access. Applications include multi-turn tool-use agents, code-execution agents, vision-language environments (the shipped geo3k recipe), LLM-as-judge reward loops, and on-policy distillation. Molt composes Ray for placement and asynchronous queues, vLLM for rollout, and NVIDIA AutoModel with FSDP2 for training, none forked. The runtime is an agent pool with vLLM engines behind a request router and a single trainable policy actor. Partial rollout pauses engines, broadcasts actor shards over NCCL, and resumes retained requests. An RL run names one Python module exporting an AgentRunner; rewards use Env (framework-owned LLM loop) or ChatAgent (user-owned