In June 2026, during the GTC conference held in Taipei, NVIDIA made a groundbreaking announcement—the official launch of NVIDIA Cosmos 3. This model stands as the world's inaugural fully open-source, multimodal physics AI foundation model. Leveraging a hybrid Transformer architecture, NVIDIA Cosmos 3 seamlessly integrates a suite of advanced capabilities, including visual reasoning, world generation, and action prediction. This integration empowers the model to natively comprehend and generate a diverse range of content, encompassing text, images, videos, ambient sounds, and motion trajectories.

One of the most remarkable achievements of NVIDIA Cosmos 3 is its ability to drastically reduce the training cycle for physics AI. What once took months can now be accomplished in a matter of days, marking a significant leap forward in AI development efficiency. In benchmark tests such as Artificial Analysis and Physics-IQ, Cosmos 3 has demonstrated its superiority by securing the top rank, showcasing its exceptional performance and accuracy.

The applications of NVIDIA Cosmos 3 span across various cutting-edge fields, with robotics and autonomous driving being prominent examples. Leading companies in these sectors, including Li Auto and AgiBot, have already embraced this revolutionary model, leveraging its capabilities to drive innovation and enhance their products and services.

In conjunction with the launch of NVIDIA Cosmos 3, NVIDIA also established the Cosmos Global Developer Collaboration Alliance. This alliance aims to foster technological iteration and facilitate the implementation of next-generation open-world models. By bringing together developers from around the globe, NVIDIA seeks to create a vibrant ecosystem where ideas can be exchanged, collaborations can flourish, and the boundaries of AI can be pushed even further.