NVIDIA has turned its Cosmos 3 world model into a set of tools that behave more like a simulator than a video generator. The four new checkpoints, all post-trained from Cosmos 3 Nano and published on Hugging Face, let teams building self-driving cars and robot arms generate sensor data and play it back in a loop with a driving or control policy. For anyone who pays for test fleets or hand-built simulation scenes, that is the reason to look.
Each checkpoint has a narrow job. Cosmos3-Nano-Transfer-Auto produces synchronised driving video for up to 11 cameras, fisheye lenses included, and can output LiDAR alongside, with a scene map setting road layout and traffic and a text prompt setting weather, light and time of day. Cosmos3-Nano-Sim-Auto is a distilled, faster version that generates short chunks step by step in a few denoising passes, so a policy can react to what it sees. Cosmos3-Nano-Sim-Transfer takes depth, segmentation, edge and blur maps from a physics simulator and renders them as realistic colour video, leaving the physics in the simulator. Cosmos3-Nano-Sim-Bimanual predicts the next frames from the actions of a two-armed robot, with support for YAM and AgiBot hardware.
The licence is OpenMDW1.1, which allows commercial use under its terms. The hardware bar is high: NVIDIA's serving setup was tested on a single GB200, and the company stresses that this is not a minimum, so smaller GPUs are unproven. Training ran in four stages, ending with the model learning from its own generated history so that small errors do not pile up over long runs. Early interest is real, with model pages showing between about 2,850 and 6,320 downloads each.
Our Cosmos 3 review describes a general world model for robots and cars; these checkpoints are the practical follow-through, aimed at closed-loop testing where a policy acts and the world responds. The comparison that matters for most teams is with conventional simulators and game-engine rendering. Cosmos does not replace them: Sim-Transfer explicitly relies on one for the physics and only supplies the realistic look, which is often where synthetic data falls down.
The caveats come partly from NVIDIA itself. It says it has not yet established where these models are faithful enough for training and evaluating policies, which is the whole point of a simulator. Errors can still compound over long generations. The blog post gives no resolution, frame rate, clip length or benchmark figures, and it lists 16B trainable parameters while the model cards say 15B. Teams sizing hardware should check the cards and the linked technical report before committing.