New NVIDIA Cosmos 3 Open Models for Interactive World Simulation

NVIDIA introduced four new open-weight checkpoints for its Cosmos 3 model family, designed to generate interactive world simulations and multi-sensor driving scenes. All models are post-trained variants of Cosmos 3 Nano, each containing 16 billion trainable parameters and using OpenMDW 1.1, which allows both commercial and non-commercial use under its licensing terms.

The models include Sim-Bimanual, which predicts visual outcomes from robotic actions, and Sim-Transfer, which converts physics simulator outputs into generated observations. For autonomous driving, the checkpoints enable synchronized camera and LiDAR generation, allowing researchers to manipulate road layouts and traffic signals to study policy responses in simulated environments.

Cosmos3-Nano-Transfer-Auto generates synchronized video across up to 11 camera views, including fisheyes, and supports joint RGB–LiDAR generation with World Scenario Map controls for scene structure and traffic motion. Developers can adapt the model to custom sensor setups or driving domains by post-training it on their data, enabling offline synthetic data generation or streaming simulation.

Cosmos3-Nano-Sim-Auto is a causally distilled streaming variant of Transfer-Auto, generating successive synchronized camera video and LiDAR chunks in a few denoising steps. Sim-Bimanual simulates how environments evolve in response to bimanual robot actions, supporting interactive two-arm simulations for specific robot embodiments controlled by policies or human input.