According to materials released by NVIDIA, Cosmos 3 is the world’s first fully open omnimodal physical AI foundation model. Built on NVIDIA’s proprietary hybrid Transformer architecture, the model integrates both reasoning and generation branches, enabling visual reasoning, world generation, and action prediction within a single model. It can accurately reproduce real-world physical rules such as gravity, collision, and friction. Unlike general multimodal models primarily designed for entertainment content generation, Cosmos 3 natively supports the input and generation of five modalities: text, images, video, ambient audio, and human and robotic motion. Pretrained on large-scale physically annotated datasets, the model can autonomously generate realistic scene data and motion trajectories that follow real-world physical laws.
Official information shows that Cosmos 3 is planned in multiple model tiers. The Super version, with larger parameters, is designed for automakers and humanoid robotics companies for large-scale data synthesis and advanced algorithm development. The Base version targets small and midsize R&D teams across industries, while the Tiny version is optimized for local deployment on edge and embedded hardware. NVIDIA has fully opened the model weights, source code, and supporting evaluation datasets, with related resources being made available on open platforms such as Hugging Face and GitHub. Global developers can access the resources for free and fine-tune them as needed. In terms of deployment efficiency, NVIDIA says the model can compress traditional simulation and debugging cycles for robotics, autonomous driving, and industrial digital twin projects from several months to just a few days, helping reduce the cost of data collection and testing for physical AI development.
Alongside the product announcement, NVIDIA also unveiled its ecosystem strategy by forming the NVIDIA Cosmos Coalition, a global collaboration alliance. The founding members include leading AI labs and robotics companies such as Agile Robots, Black Forest Labs, Runway, LTX, and Skild AI. Under the coalition’s collaboration framework, member organizations can share the full Cosmos 3 technology stack, access NVIDIA DGX Cloud computing resources as needed, and exchange algorithms, benchmarks, and real-world industry testing data. The alliance aims to jointly develop common standards for open world models and accelerate the iteration and commercialization of next-generation physical AI technologies.
Publicly available information indicates that several automakers and automation equipment manufacturers have already begun internal testing of projects based on Cosmos 3. Potential deployment scenarios include humanoid robot development, autonomous driving simulation, industrial visual inspection, and realistic digital content creation.