NVIDIA

NVIDIA has launched Cosmos 3, an open physical AI foundation model built on a mixture-of-transformers architecture. The model family targets robotics, autonomous vehicles and vision agents, with Super, Nano and upcoming Edge variants. The signal is NVIDIA extending its influence from AI compute into the model and software stack for physical AI.

•Super, Nano and Edge target accuracy, speed and edge inference

•Cosmos 3 tops open-model benchmarks; coalition members span robotics, video generation and AV

The fact

NVIDIA has launched Cosmos 3, an open physical AI foundation model built on a mixture-of-transformers architecture. It natively supports text, image, video, ambient sound and action understanding and generation. Trained on billions of multimodal physical AI samples, Cosmos 3 can reduce training and evaluation cycles from months to days. Cosmos 3 ranks first among open models across multiple physical AI benchmarks for world generation, action policy and vision understanding. The lineup includes Super, Nano and the upcoming Edge version, while the Cosmos Coalition includes Agile Robots, Runway and Skild AI.

The Assessment

Cosmos 3 extends NVIDIA beyond hardware into the software layer of physical AI development. By combining synthetic data generation, world modelling and action-policy training in an open model family, NVIDIA lowers entry costs for robotics, autonomous vehicles and vision agents. The strategic signal is not just model performance, but NVIDIA's attempt to make its stack the default path from simulation to real-world deployment — a playbook echoing CUDA's hold on GPU compute. For BTW readers, the Edge variant's real-time inference demands will test whether NVIDIA's embedded silicon (Jetson) can keep pace with latency-sensitive robotics workloads.

What to Watch

Watch Cosmos Edge latency and power draw in real deployments, coalition-built applications, and whether the model family pushes Jetson and DGX demand.