What it is for
Every robotics and autonomy team hits the same wall: real-world data is slow and expensive to collect, and the rare cases that matter most (a child stepping into the road, a box that is heavier than it looks) are the hardest to film. Cosmos is Nvidia’s answer: models that generate and transform physically plausible video so a policy can train on thousands of variations of a scene instead of the one you recorded. It is a world model in the training-data-engine sense, not a game.
From Predict, Transfer and Reason to Cosmos 3
Nvidia announced Cosmos at CES in January 2025 as a family of separate open models:
- Predict generates what happens next from text, an image or a clip.
- Transfer turns structured inputs such as segmentation, depth or raw simulator output into photorealistic video, so one simulated scene can be re-rendered in rain, at night or in a different warehouse.
- Reason is a vision language model that critiques physical plausibility and can filter generated data or label real footage.
Cosmos 3, launched at Computex in June 2026, collapses those roles into a single mixture-of-transformers omnimodel that understands and generates text, images, video, ambient sound and actions. It acts as a vision language model, as a world model that predicts future states, and as a backbone for "world action models" that output robot actions directly. It ships as Super (highest physics accuracy) and Nano (fast video and action reasoning), with Edge announced for real-time on-device use. Nvidia says it ranks first among open models on benchmarks including Physics-IQ and RoboArena.
Who uses it
The launch partner lists are long and worth reading as a map of the industry. Nvidia named a Cosmos Coalition of Agile Robots, Black Forest Labs, Generalist, LTX, Runway and Skild AI, with robotics developers including Doosan Robotics, LG Electronics and Samsung and autonomous-vehicle maker Li Auto among the adopters, and inference available through Azure, CoreWeave, Nebius, Baseten and others. The original 2025 launch counted humanoid makers such as 1X, Agility and Figure among early users.
Why Nvidia gives it away
Every synthetic training hour runs on GPUs. Giving away a good world model grows the market for the hardware it runs on, the same way CUDA did for deep learning. That does not make the model worse; it does mean its roadmap will favour workloads that need more compute, and it is optimised for Nvidia hardware first. AMD buying World Labs and Amazon backing Odyssey and Decart are the competitive responses.
Cosmos or a closed model?
- Choose Cosmos when you need to fine-tune on your own scenes, keep data in your environment, or put a world model in a production pipeline this quarter.
- Choose a closed model (Genie, Runway, Odyssey) for interactive experiences or when the vendor offers something specific, such as Waymo-style driving simulation built on Genie 3.
- Either way, measure sim-to-real transfer on your task. A clip that looks right is not evidence that a policy trained on it will work on hardware.
Related
- World models: the explainer and the full list of world models
- Google Genie vs other world models: Cosmos compared with Genie, World Labs and others
- AMI Labs and JEPA: the approach that tries to need less training data
- Humanoid robot companies: the robot makers consuming this data
Frequently asked questions
What is Nvidia Cosmos?
Cosmos is Nvidia’s family of world foundation models for physical AI, first announced at CES in January 2025. The models generate and transform video of physical scenes, predict future states and reason about them, so robotics and autonomous-vehicle developers can produce synthetic training and test data. They are released with open weights.
What is Cosmos 3?
Cosmos 3 is the June 2026 generation, launched at Computex. Nvidia calls it a fully open omnimodel: one mixture-of-transformers model that understands and generates text, images, video, ambient sound and actions. It ships in Super (highest physics accuracy) and Nano (fast) sizes on Hugging Face, with an Edge size for on-device inference announced as coming later.
What are Cosmos Predict, Transfer and Reason?
In the first Cosmos generations they were separate model families. Predict generates future video from text, image or video input. Transfer turns structured inputs such as segmentation maps, depth or simulator output into photorealistic video, which is useful for varying lighting, weather and textures. Reason is a vision language model that judges whether a scene or a generated clip is physically plausible. Cosmos 3 folds these roles into one model.
Is Nvidia Cosmos free?
The weights are free to download from Hugging Face under Nvidia’s open model license, which permits commercial use subject to its terms. Running it is not free: the larger models need data-centre GPUs, which is exactly the business Nvidia is in.
Nvidia Cosmos vs Google Genie: which should I use?
Use Cosmos when you need training or test data for a robot or vehicle and want to run the model in your own environment. Use Genie when you want an interactive world a person can explore, for games, film previz or demos. Genie is the more impressive experience; Cosmos is the more practical tool for engineering teams.
Sources: Nvidia (CES 2025 Cosmos launch; Computex 2026 Cosmos 3 launch), GamesBeat, Hugging Face. Benchmark rankings are as reported by Nvidia.