NVIDIA launches Cosmos 3, an open-source multimodal model for physical AI

NVIDIA has unveiled Cosmos 3, an open-source multimodal model for physical AI. It uses a hybrid Transformer architecture to understand and generate text, images, video, sound, and actions with top-tier physics simulation accuracy. This can cut training time from months to days.

NVIDIA also launched the Cosmos Coalition, teaming up with Agile Robots, Black Forest Labs, Runway, and others to advance world model tech. CEO Jensen Huang said breakthroughs in multimodal reasoning and vision models are speeding up the arrival of the physical AI era.

Cosmos 3 comes in several versions: Super for training robots and self-driving cars, Nano for quick video and action analysis, and Edge coming soon for real-time use on devices. It tops open-source benchmarks in world generation, action planning, and visual understanding.