NVIDIA Unveils Fully Open-Source Physics AI Model Cosmos3, Enabling Multimodal Generation and Drastically Shortening Training Time

On June 1, NVIDIA launched Cosmos3, an open-world foundation model designed for physical artificial intelligence (AI). Built on a hybrid Transformer architecture that combines a reasoning Transformer with specialized generative Transformers, the model is trained on a multimodal physical AI dataset comprising billions of samples of text, images, videos, audio effects, and motion trajectories. Cosmos3 natively supports understanding and generation of text, images, video, environmental audio, and motion content. It delivers industry-leading physical simulation accuracy, reducing the training and evaluation cycle for physical AI from months to just days. On mainstream physical AI benchmarks, Cosmos3 ranks first in world-generation fidelity, action-policy capability, and visual understanding. The model is available in multiple versions: Cosmos3 Super is optimized for fine-tuning robotics and autonomous driving models; Cosmos3 Nano enables high-quality video parsing and motion reasoning in seconds; and Cosmos3 Edge, designed for real-time inference at the edge, will be released soon. NVIDIA also announced the formation of the Cosmos Coalition, bringing together global world-model research teams and AI developers to advance next-generation world-model technologies. Developers can use Cosmos3 as a backbone network for multimodal vision-language foundation models, world models/video foundation models, or world-action models.