NVIDIA Cosmos is a family of open “world foundation models”: AI models trained to understand and generate video of the physical world, so robot and self-driving developers can create training data and test scenarios without filming everything in reality. The current generation, Cosmos 3, was announced on June 1, 2026, and combines what used to be separate models into one that reads and generates text, images, video, sound and robot actions (NVIDIA press release). The weights are free to download from Hugging Face.
What Cosmos is used for
Physical AI systems need huge amounts of video showing how the world behaves: a forklift backing up, a cup sliding off a table, a car meeting a cyclist at night. Recording all of that is slow and some of it is dangerous. NVIDIA positions Cosmos for four jobs, per the Cosmos product page:
- Synthetic data: generating varied video to train perception and robot policies.
- Closed-loop testing: simulating what happens next so a robot or driving stack can be evaluated before real-world trials.
- Reasoning about scenes: answering questions about what’s happening in a video, as a vision-language model.
- A base for robot models: developers post-train it into models that output actions.
If the idea of a world model is new, our Korean explainer what a world model is compares Cosmos with Google’s Genie and Meta’s V-JEPA.
What changed in Cosmos 3
Earlier Cosmos releases split the work across separate model lines, and NVIDIA’s docs still list them (Cosmos documentation):
- Cosmos Predict: generates future video from text, an image or a video clip.
- Cosmos Transfer: turns simulator output or control inputs, such as segmentation maps or LiDAR, into photorealistic, varied video.
- Cosmos Reason: a vision-language model that reasons about physical scenes.
Cosmos 3 merges understanding and generation in one model. It pairs an autoregressive “reasoning tower” for text with a diffusion “generation tower” for images, video and actions, a design NVIDIA calls Mixture-of-Transformers (Cosmos3-Super model card). NVIDIA hasn’t said Predict, Transfer and Reason are discontinued.
Cosmos 3 sizes
- Cosmos3-Super, 64B parameters: the most accurate version, aimed at post-training robot and autonomous-driving models. Released May 31, 2026.
- Cosmos3-Nano, 16B parameters: faster video and action reasoning.
- Cosmos 3 Edge, 4B parameters: announced in July 2026 for local use; NVIDIA described it on its Hugging Face blog.
NVIDIA’s NGC catalog also lists task-specific Cosmos 3 variants, including text-to-image, image-to-video and a robot policy model trained on the DROID dataset.
License: free for commercial use
Cosmos 3 is released under the Linux Foundation’s OpenMDW-1.1 license, which allows commercial and non-commercial use, according to the model cards. That’s a change from earlier models such as Cosmos-Predict2.5, which use the NVIDIA Open Model License. The NIM containers NVIDIA ships for deployment have their own NVIDIA software license, so read that separately if you deploy through NIM.
What you need to run it
The model cards set the bar at data-center GPUs:
- NVIDIA Ampere, Hopper or Blackwell GPUs, on Linux, in BF16 precision. FP8 and FP4 aren’t officially supported yet.
- NVIDIA’s example setup for Cosmos3-Super is eight H100, H200 or A100 GPUs, or two H200s. Nano examples use a single H200.
- NVIDIA doesn’t publish a minimum VRAM figure for Cosmos 3. For scale, the older Cosmos-Predict2.5-2B needs about 32.5 GB of VRAM for 720p video generation (model card).
Without that hardware, you can try Cosmos in the browser at build.nvidia.com or through cloud providers NVIDIA names, including Baseten, CoreWeave, Azure and Nebius.
How to get started
- Read the model card for the size you need on Hugging Face (nvidia/Cosmos3-Super or Cosmos3-Nano).
- Get the code from NVIDIA’s Cosmos GitHub repository.
- Pick a runtime. NVIDIA documents vLLM, vLLM-Omni, SGLang and Hugging Face Diffusers.
- Follow a task recipe in the Cosmos Cookbook, such as generating synthetic driving scenes or post-training for a robot.
Who uses it
NVIDIA lists Agile Robots, Black Forest Labs, Generalist, LTX, Runway and Skild AI as founding members of its Cosmos Coalition, and names Doosan Robotics, LG Electronics, Samsung and Li Auto among companies building on Cosmos 3. These are NVIDIA’s announcements; most companies haven’t published results from using it. LG’s plan is covered in our article on LG’s robotics data factory with Cosmos and Isaac.