NVIDIA's Cosmos 3 open model supercharges physical AI with 5x faster training
First fully open omnimodel that reasons and generates across vision, video, sound, and action.
NVIDIA today unveiled Cosmos 3, an open world foundation model designed specifically for physical AI — robots, autonomous vehicles, and vision agents that need to understand and act in the real world. The model uses a novel mixture-of-transformers architecture that pairs a reasoning transformer with an expert generation transformer, enabling it to grasp object interactions, motion, and spatial-temporal relationships before generating video or action trajectories. Trained on billions of multimodal samples spanning text, images, video, ambient sound, and action data, Cosmos 3 delivers state-of-the-art performance across multiple leaderboards including Artificial Analysis, Physics-IQ, and RoboLab. The lineup includes Cosmos 3 Super for maximum accuracy, Cosmos 3 Nano for near-real-time reasoning, and an upcoming Cosmos 3 Edge for on-device inference.
Beyond the model itself, NVIDIA launched the Cosmos Coalition — a global collaboration with leading AI labs and robotics companies such as Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI. The coalition aims to accelerate open world model development by sharing research, evaluation techniques, and access to NVIDIA DGX Cloud for large-scale training. According to CEO Jensen Huang, this breakthrough marks “the big bang of physical AI,” giving developers a generational leap in building systems that perceive, reason, plan, and act in the physical world. Cosmos 3 is available now as an open model, with the expected impact of reducing physical AI development cycles from months to days.
- Cosmos 3 is the first fully open omnimodel handling text, images, video, sound, and action natively.
- Its mixture-of-transformers architecture combines reasoning and generation for world simulation and action prediction.
- The Cosmos Coalition includes partners like Runway, Skild AI, and Black Forest Labs to drive open physical AI development.
Why It Matters
Accelerates robot and AV development from months to days, making physical AI accessible to more teams.