NVIDIA on May 31 launched Cosmos 3, an open world foundation model family for physical AI that the company says unifies vision reasoning, world generation and action prediction. Two variants—Cosmos 3 Super and Cosmos 3 Nano—became available immediately; a third, Cosmos 3 Edge, is “coming soon” for real-time edge inference without a date. The release, at GTC Taipei/COMPUTEX 2026, extends the chipmaker’s push into software and models.
The move positions a leading AI infrastructure company as a supplier of foundational models for robotics, autonomous vehicles and industrial automation. By releasing the models under a permissive license with open code and tools, NVIDIA aims to make its physical AI stack a default layer for commercial developers.
NVIDIA said Cosmos 3 uses a Mixture-of-Transformers architecture with separate reasoner and generator components, capable of understanding and generating text, images, video, ambient sound and actions. The Nano variant pairs an 8-billion-parameter reasoner with an 8-billion-parameter generator; Super scales each to 32 billion parameters. The training mix includes video, image, language and action data, though full provenance was not disclosed.
The two available models are accessible via Hugging Face, GitHub, build.nvidia.com and as NIM microservices, under the OpenMDW 1.1 license hosted by the Linux Foundation. NVIDIA says the license permits commercial and non‑commercial use, training, modification, redistribution and deployment. The Edge model has no public hardware support details or timeline.
NVIDIA claimed that Cosmos 3 ranks first among open models on physical AI benchmarks including Artificial Analysis, Physics-IQ and RoboArena, and that it can reduce training and evaluation cycles “from months to days.” Both claims originate from the company and lack independent third‑party verification at the time of announcement.
The company identified intended applications in robotics, autonomous vehicles, smart spaces, industrial vision, simulation and synthetic data generation.
A new Cosmos Coalition, with founding members Agile Robots, Black Forest Labs, Generalist, LTX, Runway and Skild AI, will promote an open ecosystem, though no specific joint projects were disclosed.
Cosmos 3 arrives 17 months after the original Cosmos platform launched in January 2025 and follows a March 2026 preview. It represents a unified family combining reasoning and generation, a step beyond the earlier, more fragmented offerings.
The Edge variant’s release remains undefined, all benchmark leadership rests on NVIDIA’s own testing, and the training data’s copyright and licensing status are unresolved—factors that enterprise adopters will need to investigate. Still, the immediate availability of open-weights models under a clear commercial license puts a concrete option into developers’ hands, raising the pressure on rivals to match openness and independent validation.