NVIDIA launches Cosmos 3, an open world foundation model for physical AI

NVIDIA/Cosmos 3

NVIDIA announced Cosmos 3, an open multimodal world model trained on 20 trillion tokens to understand text, images, video, sound and actions. Available in Super (32B) and Nano (8B), it is designed for robotics training and autonomous-vehicle simulation. Built by a coalition including Runway, Black Forest Labs and Skild AI.

• Multimodal training accelerates physical AI development by providing a shared model of physics.
• Open model via coalition reduces NVIDIA's development cost and speeds adoption.
• Can serve as a vision-language model, world model for video prediction, or backbone for robot training.
• Edge variant coming for real-time inference on robot platforms.