NVIDIA Releases Cosmos 3 World Model for Robotics and Autonomous Driving
Taylor Wilson
Nvidia open-sources Cosmos 3, a world model trained on 20 trillion tokens to help machines understand real-world physics — another move in its expansion from chips to AI platforms.
What exactly is a "world model"?
A regular AI watches video and learns what scenes look like. A world model goes one layer deeper — how objects move and what happens when they collide.
In plain terms = it gives machines a kind of "physical intuition," letting robots and self-driving cars anticipate what comes next.
Cosmos 3 was trained on nearly 1 billion images and 400 million video clips, plus environmental audio and robotic action data.
Action data — what sets Cosmos apart from ordinary video generators?
Ming-Yu Liu, VP of Nvidia's Cosmos Lab, put it directly: "Its goal is to model how machines move, not just how scenes look."
This means → Cosmos 3 can output a robot's joint angles, gripper positions, and motion trajectories — not just a polished video.
It can also simulate collisions, rare road incidents, and other dangerous scenarios that are nearly impossible to capture repeatedly in the real world.
Three versions — who is each one for?
Super: highest physical accuracy, built for robotics training and autonomous driving where realism is non-negotiable.
Nano: generates results in milliseconds, suited to applications that need near-instant response.
Edge (coming soon): runs locally on-device, no cloud compute required.
This reflects Nvidia's aim to cover the full stack from cloud to edge — not just ship one big model.
Why go open-source?
Open-sourcing lets hardware makers customize the model to their own needs — the same logic behind Nvidia's earlier open-source Nemotron series.
Nvidia also launched a partner alliance with founding members Agile Robots, Black Forest Labs, and Runway.
In plain terms = open source is a "get everyone using it first" play. The wider the adoption, the deeper the ecosystem lock-in — and the stronger the moat for selling chips and platforms later.
Who else is in this race?
Fei-Fei Li's World Labs and Yann LeCun's AMI Labs are both building world models; the space is heating up.
Nvidia's difference: it is not a startup but a giant expanding from chips into AI models and software platforms.
This means → Cosmos 3 is not a standalone product. It is another piece of Nvidia's "chips + platform + ecosystem" strategic puzzle.
Content is for reference only, not financial advice.