NVIDIA Launches Cosmos 3, the First Open Omnimodel for Physical AI
Summary: NVIDIA unveiled Cosmos 3 at Computex 2026 — the industry's first fully open omnimodel that natively understands and generates text, image, video, ambient sound, and action trajectories in a single system.
Key Facts
- Built on a mixture-of-transformers architecture pairing a reasoning transformer with an expert generation transformer, enabling physics-grounded simulation before action
- First fully open omnimodel: supports all five modalities natively; weights available as Cosmos3-Super and Cosmos3-Nano on Hugging Face
- Compresses robot and autonomous-vehicle training cycles from months to days via high-quality synthetic data generation
- NVIDIA launched the Cosmos Coalition — Agile Robots, Black Forest Labs, Runway, Skild AI and others — to advance open world models
Why It Matters
Physical AI has been bottlenecked by the cost and slowness of real-world data collection. Cosmos 3 gives robotics and AV startups a foundation model they can fine-tune without building their own simulation stack — sharply lowering the barrier to embodied AI development and accelerating the path from lab to real-world deployment.
Read More
- NVIDIA Cosmos 3 Press Release — NVIDIA Newsroom
- Model Overview on Hugging Face — Hugging Face Blog