NVIDIA releases Cosmos 3 Nano simulation checkpoints
NVIDIA has released open Cosmos 3 Nano checkpoints for driving, robotics and surgical simulation, plus new reinforcement learning pipelines and agent skills in Cosmos Framework, to help physical-AI developers adapt world models and deploy them.
NVIDIA detailed in a technical blog post new tooling for its Cosmos world-model platform. The release includes Cosmos 3 Nano simulation recipes and example checkpoints for autonomous vehicle, robot and surgical scenarios. The checkpoints run on PyTorch and vLLM-Omni and were tested on Hopper and Blackwell GPUs.
One checkpoint, Cosmos3-Nano-Transfer-Auto, generates synchronized multi-camera video and LiDAR. It is conditioned on a text prompt and World Scenario Map controls covering lanes, objects, signs and road boundaries. It denoises a whole multi-view clip at once for offline synthetic data generation, producing up to 400 frames at 30 fps in 480p or 720p. Longer sequences use autoregressive chunking. Changing only the prompt can turn a daytime drive into a rainy night while the road layout and vehicle trajectories stay the same. In NVIDIA's four-GPU reference run, the model produced seven 720p videos of 297 frames each, plus 99 LiDAR sweeps.
Cosmos Framework, NVIDIA's open, modular framework for training and inference, now includes online and offline reinforcement learning pipelines: DPO, GRPO with vLLM rollouts, PPO (including for embodied policies in simulation), on-policy distillation, and DiffusionNFT for diffusion world models. NVIDIA says the framework is expanding beyond Cosmos 3 to support Alpamayo, GR00T, Qwen, pi0.5 and other compatible Hugging Face models. It also ships agent skills for setup, inference, post-training and troubleshooting.
For deployment, Cosmos NIM microservices provide GPU-accelerated serving. The post explains how to choose a NIM configuration for a given GPU and send inference requests. The source material is cut off before it gives full deployment details or any benchmarks.