Rongchai Wang
Jun 01, 2026 06:02
NVIDIA launches Cosmos 3, a cutting-edge model for physical AI, enabling faster development in robotics and autonomous systems.
NVIDIA has officially launched Cosmos 3, its latest open world foundation model designed to accelerate advancements in physical AI. Built on a novel mixture-of-transformers architecture, Cosmos 3 integrates vision reasoning, synthetic world generation, and action prediction into a unified system. The announcement, made at NVIDIA GTC Taipei on June 1, 2026, signals a major leap forward in AI-driven robotics and autonomous vehicle (AV) development.
Why Cosmos 3 Matters
Cosmos 3 is the first fully open “omnimodel” capable of multimodal generation across text, image, video, ambient sound, and action. By significantly reducing training cycles, it allows developers to generalize physical AI solutions with less data and lower costs. This capability is crucial for creating robots and AVs that can perceive, reason, and act in real-world environments without extensive retraining.
According to NVIDIA CEO Jensen Huang, “The big bang of physical AI is just around the corner. Cosmos 3 offers a generational leap for building systems that can reason, plan, and act in the physical world.”
Technical Innovations
The model introduces a new mixture-of-transformers architecture that combines reasoning and generation transformers. This setup enables Cosmos 3 to understand object interactions, spatial relationships, and motion before generating predictive video and action sequences. Trained on billions of multimodal samples, the system provides a robust foundation for physical AI policy development and synthetic data generation.
Cosmos 3 has already topped benchmarks across key physical AI evaluations, including Physics-IQ, PAI-Bench, and RoboArena, solidifying its position as a leader in the space.
Expanding the Ecosystem: The Cosmos Coalition
Alongside the model launch, NVIDIA introduced the Cosmos Coalition, a collaborative initiative with AI labs and robotics firms like Agile Robots, Black Forest Labs, and Skild AI. The coalition aims to drive open development of world models, fostering interoperability and innovation across industries.
Cosmos 3 integrates seamlessly with NVIDIA’s existing AI stack, including Omniverse for 3D simulation, DGX Cloud for large-scale training, and NeMo Curator for video processing. These integrations make it a cornerstone of NVIDIA’s broader physical AI strategy, which includes applications in robotics, autonomous driving, and industrial automation.
Market Implications
For investors and industry stakeholders, Cosmos 3 underscores NVIDIA’s dominant position in the AI hardware and software market. The platform’s focus on physical AI aligns with growing demand for robotics and AV solutions, sectors projected to see explosive growth over the next decade. NVIDIA’s ability to provide both the hardware (data center GPUs) and foundational models positions it as a critical player in this emerging arena.
Cosmos 3’s open-access model, available through platforms like Hugging Face and NVIDIA’s NGC, also enhances its appeal to developers, potentially accelerating adoption across industries. Early use cases include robotics applications from partners like Boston Dynamics and Li Auto, as well as industrial AI solutions in smart manufacturing and warehouse safety.
Availability
Cosmos 3 Super and Nano models are available now, with the Edge variant coming soon for real-time edge inference. Developers can access the models via Hugging Face, GitHub, and NVIDIA’s build portal. NVIDIA also offers cloud-based training and deployment options through partners like Microsoft Azure and CoreWeave.
With Cosmos 3, NVIDIA continues to push the boundaries of AI innovation, setting the stage for transformative advancements in robotics and autonomous systems. As the physical AI ecosystem matures, NVIDIA’s leadership in both model development and hardware infrastructure positions it at the forefront of this rapidly evolving field.
Image source: Shutterstock
Be the first to comment