NVIDIA Cosmos 3 is here — the world’s first fully open omni-model for physical AI, combining vision reasoning, world generation, and action prediction in a single foundation.
Among open models, Comsos 3 is ranking first on more than eight leaderboards for vision reasoning, text to image, image to world and world action generation with state-of-the-art physics accuracy.
Built on a breakthrough mixture-of-transformers architecture, Cosmos 3 pairs an autoregressive reasoner with a diffusion-based generator to ground outputs in real physical understanding — how scenes evolve, objects move, and actions change an environment.
One model checkpoint to simplify and accelerate physical AI training.
Explore the full release:
✅ Cosmos 3 Nano and Super checkpoints
✅ Open code and post-training recipes
✅ Open datasets for robotics, autonomous driving, physics simulation, human motion and warehouse operations
Get started:
📖 Read the Technical Blog
🤗 Read the Hugging Face Blog
📥 Download Cosmos 3
🐙 Customize Models with Cosmos 3
🔬 Try Cosmos 3 on build.nvidia.com