Nvidia on Monday introduced Cosmos 3, an open world model designed to help robots, autonomous vehicles and other physical systems better understand, simulate and navigate real-world environments.
The model is designed to generate and predict actions in physical settings rather than simply creating images or videos. Nvidia said Cosmos 3 was trained on 20 trillion tokens of multimodal data, including nearly 1 billion images, 400 million real and synthetic videos, ambient audio, text and action data collected from humans and robots. According to Axios, the inclusion of action data enables the model to simulate how machines move and interact with the physical world rather than only modeling visual scenes.
Nvidia said Cosmos 3 combines vision reasoning, world generation and action prediction in a single system. The company described it as an open foundation model for physical AI that can process and generate text, images, video, sound and action data. Its developers can use the platform to create task-specific models for robotics, autonomous vehicles and other applications involving physical environments.
The company released two versions immediately. Cosmos 3 Super is aimed at applications requiring higher physics accuracy, including robotics and self-driving vehicle training, while Cosmos 3 Nano is optimized for faster inference and can generate results in fractions of a second. Nvidia also said a Cosmos 3 Edge model capable of running locally on edge devices is under development. Nvidia's announcement outlined the three-tier approach as part of a broader strategy to support different deployment environments.
Ming-Yu Liu, vice president of Nvidia's Cosmos Lab, told the outlet that autonomous actions are a key distinction between Cosmos 3 and conventional video-generation systems. The model can generate data such as robot joint angles, gripper positions and movement trajectories, allowing developers to train machines to manipulate objects and navigate physical spaces.
Nvidia also announced the formation of the Cosmos Coalition, a group of companies supporting open world model development. Founding participants include Agile Robots, Black Forest Labs, Generalist, LTX, Runway and Skild AI, according to the company. Nvidia said coalition members will contribute models, research and evaluation methods while using Cosmos technologies and Nvidia infrastructure.
The launch builds on Nvidia's broader push into physical AI. At CES 2025, the company introduced the original Cosmos world foundation model platform, which was designed to generate synthetic data for robotics and autonomous vehicle development. At the time, Nvidia said the platform would help reduce the cost and complexity of collecting large volumes of real-world training data, according to Reuters and Nvidia's CES announcement.
The announcement comes as interest in world models grows across the technology sector. World models are designed to simulate and predict how real-world environments behave, allowing machines to test actions and outcomes in virtual settings before operating in physical environments.
Physical AI technologies have also attracted attention from defense organizations and military contractors that increasingly rely on autonomous systems, simulation platforms and robotics for training, logistics and operational planning. The ability to generate rare or dangerous scenarios, including vehicle incidents and robot collisions, is among the capabilities Nvidia highlighted for Cosmos 3. The company said such events are often difficult, costly or unsafe to capture repeatedly in the real world.
Liu told Axios that the goal of world models is to help physical agents become more adaptable by understanding how the world works and using that understanding to plan actions.
