NVIDIA launched Cosmos 3 on June 1, 2026, at its GTC Taipei event during Computex. The model is free, open, and available to download right now. It is built for one audience: any organization building or deploying machines that need to understand and act in the physical world.
The central challenge in physical AI is data. Unlike a chatbot, which can be trained on text from the internet, a robot or autonomous vehicle needs to learn from physical scenes: what happens when an object falls, how a forklift moves through a corridor, what a pedestrian looks like stepping off a pavement in the rain. Collecting that data in the real world is expensive, slow, and sometimes dangerous. Generating it artificially, with enough physical accuracy for a machine to actually learn from it, has been the core bottleneck.
Cosmos 3 is NVIDIA's answer to that bottleneck. It can generate realistic video of scenarios that are difficult or impossible to film: a warehouse collision, a surgical robot performing a procedure, a self-driving car encountering road debris. Developers feed in a text description or an image, and the model produces video that obeys real-world physics closely enough to be used as training material.
What is new in Cosmos 3 compared to earlier NVIDIA releases is consolidation. Previous versions required developers to use separate tools for scene understanding, video generation, and robot action planning. Cosmos 3 combines all of that into one model. The system first reasons about what is happening in a scene, then generates either video or specific physical instructions for a robot, such as joint movements and grip positions, without switching between tools.
Two versions are available. The smaller one, called Cosmos 3 Nano, runs on a single high-end workstation graphics card. The larger one, called Cosmos 3 Super, is designed for data center use and large-scale data generation. Both are free to download, with open training scripts on GitHub for organizations that want to adapt the model to their own environments.
Adoption is already wide across heavy industries. Samsung, LG Electronics, and Doosan Robotics are using Cosmos for robotics. Li Auto is using it for autonomous vehicles. In healthcare, Johnson & Johnson MedTech and CMR Surgical are using Cosmos-based workflows to train and validate surgical robots before clinical deployment. That last application is notable: a surgical robot trained partly on AI-generated physical simulations is now a real commercial product path.
NVIDIA also launched a new industry group alongside Cosmos 3, called the Cosmos Coalition. Founding members include Agile Robots, Runway, Skild AI, Black Forest Labs, and others, with a stated aim of contributing back to the open model rather than each company building competing closed alternatives.
The strategic logic behind NVIDIA making this free is straightforward. Every company that trains a physical AI system using Cosmos needs NVIDIA hardware to run that training. The model is the entry point; the GPU bills come later. NVIDIA has done this before with software tools, and it works.
For business operators in manufacturing, logistics, healthcare, and transportation, the practical signal is this: the cost and technical barrier to simulating your own physical environment is falling. Organizations that would never have been able to generate robotics training data at scale now have access to the same tools as the largest robotics companies in the world.