What is NVIDIA Cosmos?
NVIDIA Cosmos is a platform of open world foundation models for physical AI. It generates physics-grounded video and synthetic data, reasons about real-world scenes and supports robot policy learning. Robotics, autonomous vehicle and industrial vision developers download the models from Hugging Face under the OpenMDW 1.1 license or try them on build.nvidia.com.
Training robots and self-driving systems needs huge amounts of varied real-world data, which is slow and costly to collect. Cosmos generates plausible futures from text, image, video, sound and action inputs, so teams can create synthetic training data and test behavior in simulation. The latest Cosmos 3 is described as an omni-model that handles reasoning, world generation and action generation, with post-training scripts on GitHub for your own domain.
Main Features
- World simulation that predicts physics-grounded outcomes for closed-loop testing.
- Synthetic data generation from text, image, video, ambient sound and action inputs.
- Vision reasoning over objects, interactions and intent for inspection and autonomy.
- World Action Models that adapt pre-learned actions to speed up robot learning.
- Open models on Hugging Face under the OpenMDW 1.1 license from the Linux Foundation.
- Post-training scripts and tooling on GitHub.
- Hosted models to try on build.nvidia.com and recipes in the Cosmos Cookbook.
How to Get Started with NVIDIA Cosmos
- Visit the NVIDIA Cosmos page and choose a model for your task.
- Try a hosted model on build.nvidia.com to test prompts and outputs.
- Download the weights from Hugging Face and the scripts from GitHub.
- Post-train on your own data using a Cosmos Cookbook recipe.
NVIDIA Cosmos Pricing
The Cosmos models are openly released under the OpenMDW 1.1 license and free to download. You need your own GPU hardware or cloud compute to run and post-train them.
Who Should Use It?
- Robotics engineers generating training data
- Autonomous vehicle teams testing driving scenarios
- Industrial vision developers building inspection systems
- AI researchers working on world models
NVIDIA Cosmos Alternatives
- Genmo AI: consumer video generation with a world-model focus
- Meta Sam 2: segments objects in video rather than generating it
- Wan AI: creative text-to-video for content, not robotics
Frequently Asked Questions
Is NVIDIA Cosmos free?
The models are free to download and use under the OpenMDW 1.1 license from the Linux Foundation. Running them takes significant GPU compute, so real costs come from your own hardware or cloud instances.
What is NVIDIA Cosmos used for?
It is used to build physical AI: generating synthetic video data, simulating outcomes for robots and vehicles, reasoning about real-world scenes, and speeding up robot policy learning. It is not aimed at consumer video creation.
Where can I download NVIDIA Cosmos?
Weights are on Hugging Face, post-training scripts are on GitHub, and hosted versions can be tried on build.nvidia.com. NVIDIA also publishes a Cosmos Cookbook with domain-specific recipes.