NVIDIA
NVIDIA Cosmos
Open models for physical simulation

Key facts
- FourCosmos 3 Nano examples
- Specialist models
- 16Bper specialist Nano model
- Parameters
- OpenMDW1.1current specialist models
- Licence
- 20 Oct 2026checkpoint release
- Transfer-Auto
- Up to 11Transfer-Auto driving scenes
- Camera views
NVIDIA Cosmos generates simulated environments for robots and autonomous vehicles. Cosmos 3 Nano includes specialist driving and robot models.
What it is
NVIDIA Cosmos is a family of AI models that generate simulated environments for robots and autonomous vehicles. Teams can download the model weights and adapt them to their own environments. Generated scenes provide extra training and evaluation data, including situations that are difficult to collect with real vehicles.
Four specialist models for driving and robotics
NVIDIA’s 5 November explanation describes four Cosmos 3 Nano models with 16 billion trainable parameters each. Transfer-Auto produces driving scenes for batch data generation. Sim-Auto produces successive camera and LiDAR observations for interactive driving simulations. Sim-Transfer turns a simulator’s spatial controls into video. Sim-Bimanual predicts visual changes from a two-armed robot’s actions.
Sim-Transfer relies on the underlying simulator for physical state. Teams need to evaluate how faithfully generated observations represent their intended environment before using them to train or assess a robot.
Driving scenes with several sensors
Cosmos3-Nano-Transfer-Auto, released on 20 October 2026, can generate synchronised video from up to 11 camera views alongside LiDAR observations. Scene controls describe road structure and traffic, while text specifies appearance such as weather and lighting. Adapting it to a different sensor arrangement requires further training.
Running Cosmos locally
The specialist models use OpenMDW1.1, which permits commercial and non-commercial use under its terms. Hardware needs depend on the checkpoint, resolution, sensor configuration and runtime. NVIDIA’s model cards provide configuration-specific performance tables; use those when sizing a deployment. The earlier 7B and 14B Predict models have their own requirements and licences.
Cosmos fits teams building synthetic data or integrating generated observations into their own simulations. The world models hub compares it with other approaches.
More in World Models
All World Models →- Google DeepMindGenie 4explorable AI worlds you steer by typing, with their own sound
- RekaReka inverse dynamics modelestimating camera movement from video
- RekaReka Rho-1research across text, vision and actions
- OdysseyOdyssey-3one pretrained world model behind robots, cars, drones and games
- Google DeepMindGenie 3the reference interactive world model
- RunwayRunway GWM-1the most general commercial family