NVIDIA Cosmos 3 tops benchmarks as open physical AI models go mainstream

NVIDIA is putting real weight behind its call for open standards in artificial intelligence, and its newest move shows how far that push now reaches into robotics, self-driving cars and machine vision. In July, the chipmaker joined more than 200 companies and organizations in signing an open letter titled “Open Weights and American AI Leadership,” arguing that AI leadership should be judged by how widely open systems spread across every sector, not by a single flagship model. Building on that commitment, NVIDIA has now introduced Cosmos 3, a new family of open physical AI models designed to help machines understand and predict what happens in the real world, not just describe what a camera sees.

Key takeaways

• NVIDIA joined more than 200 companies and organizations behind an open letter pushing open AI ecosystems forward.
• NVIDIA Cosmos 3 is a new open physical AI foundation model built on a mixture-of-transformers architecture.
• The Cosmos 3 family spans three sizes: Super (64B), Nano (16B) and Edge (4B), each aimed at different deployment needs.
• Cosmos 3 ranks No. 1 on multiple independent benchmarks, including Artificial Analysis, PAIBench, Physics-IQ, RoboLab and VANTAGE-Bench.
• Companies in robotics, autonomous vehicles and vision AI, along with the expanding NVIDIA Cosmos Coalition, are already building on the model family.

NVIDIA Joins the Push for Open Physical AI Ecosystems

Open access to model weights is what lets developers actually build on top of foundation models instead of just renting access to them. NVIDIA’s decision to co-sign the “Open Weights and American AI Leadership” letter reflects a broader argument gaining traction across the industry: that ciò che consente a chiunque di scaricare, esaminare, personalizzare ed eseguire modelli aperti sulla propria infrastruttura è widespread AI adoption possible in the first place.

That principle matters most in fields where a single generic model can’t do the job. Physical AI, the branch of artificial intelligence dealing with robots, autonomous vehicles and vision systems operating in real environments, is exactly that kind of field. Every deployment is effectively a specialization problem, since a warehouse robot, a delivery vehicle and a factory camera all need to reason about entirely different physical conditions. Open world models give engineering teams the flexibility to adapt a shared foundation to each of those very different settings, rather than starting from scratch every time.

What Sets Physical AI Apart From Other AI Systems

Physical AI has to understand and predict consequences, not just recognize appearances. That distinction is the whole reason world models exist: they’re trained to learn how physical environments behave, anticipate what might happen next, and figure out which actions make sense given that prediction.

In practice, this means a world model can generate physically grounded data about how objects move and interact, simulate future states of an environment, and hand developers a foundation they can then fine-tune for a specific robot, self-driving car or vision system. Instead of collecting massive amounts of real-world footage for every new use case, teams can lean on world models simulation to generate synthetic training data and test policies before anything touches the real world. NVIDIA’s Omniverse libraries, part of its Agent Toolkit, extend this further by providing prebuilt tools for building simulation-ready environments where physical AI systems can be trained, tested and validated before real-world deployment.

Inside NVIDIA Cosmos 3, the New Open Physical AI Foundation Model

Cosmos 3 is NVIDIA’s answer to the fragmentation problem that has long slowed down physical AI development. Rather than maintaining separate models for scene understanding, data generation and action prediction, developers can now rely on a single open model family that handles all three. NVIDIA describes Cosmos 3 as a frontier open physical AI foundation omni-model built on a mixture-of-transformers architecture, combining vision reasoning, world generation and action prediction in one system.

That flexibility is the point. Developers can use Cosmos 3 as a vision language model to interpret a scene, as a physics-grounded world simulator that predicts future states and generates large-scale synthetic training data, or as the backbone for specialized world action models built for a particular robot or vehicle.

Model Variants for Every Deployment Need

The Cosmos 3 family comes in three sizes, each targeting a different balance between capability and hardware footprint:

• Cosmos 3 Super (64B) is built for high-fidelity world modeling.
• Cosmos 3 Nano (16B) is tuned for efficient reasoning and post-training.
• Cosmos 3 Edge (4B) is designed for on-device vision reasoning and robot policy deployment.

Cosmos 3 Edge is light enough to run on edge GPUs, and NVIDIA says it can be deployed across NVIDIA RTX GPUs, NVIDIA DGX systems and NVIDIA Jetson hardware, including the Jetson Thor platform. That range matters for companies that need to run inference directly on a robot or vehicle rather than routing everything through a data center.

Benchmark Leadership Across Tasks

Across independent benchmark evaluations, Cosmos 3 currently ranks No. 1 on Artificial Analysis for open-weight generazione da testo a immagine e da immagine a video, su PAIBench per la generazione di mondi, e nella categoria da immagine a video di Physics-IQ. Per la politica dei robot tasks specifically, it tops the RoboLab benchmark. Cosmos 3 Super is also the highest-ranked open model on VANTAGE-Bench for vision understanding, according to NVIDIA.

These rankings position Cosmos 3 as a genuine reference point for anyone building open physical AI models rather than a niche research release. NVIDIA’s broader physical AI stack rounds things out with Isaac GR00T for robotics, Alpamayo for autonomous vehicles and Metropolis for vision AI, meaning Cosmos 3 doesn’t operate in isolation but feeds into a wider set of tools already tailored to specific industries.

Industry Adoption and the NVIDIA Cosmos Coalition

Companies across robotics, autonomous vehicles and vision AI are already building on NVIDIA Cosmos, which suggests the model family is moving beyond a demo stage into active development pipelines. In robotics, Doosan Robotics, LG Electronics, Samsung Electronics and Skild AI are among the names NVIDIA lists as adopters. In autonomous vehicles, Li Auto, Xiaomi and Afari are working with the platform. For vision AI applications powering industrial monitoring and smart-space systems, Centific, Fogsphere, Linker Vision, Milestone Systems and Yuan are named as developers building on Cosmos.

Why does that spread across three very different industries matter? It’s a sign that a single open foundation model can genuinely serve as shared infrastructure rather than something that only works for one narrow use case — which is precisely the argument NVIDIA and its fellow signatories made in the open letter earlier this year.

The NVIDIA Cosmos Coalition is the mechanism meant to keep that momentum going. It brings riunendo costruttori di modelli mondiali, sviluppatori di IA e leader dell’IA fisica affinché forniscano modelli, ricerche e metodologie di valutazione back into the open ecosystem. NVIDIA recently expanded the coalition into Japan, where robotics and manufacturing leaders intend to join and develop open world models tailored to factories, logistics, agriculture, construction, healthcare and transportation. Combined with the industry adoption already underway, that expansion signals an open model family is becoming a shared foundation for physical AI rather than a single company’s proprietary bet.

FAQ

What are open world models and why are they important?

Open world models can be downloaded, inspected, modified and run on any infrastructure, which gives developers the customization needed to specialize physical AI applications for a specific robot, vehicle or vision system.

What distinguishes physical AI from other AI types?

Physical AI requires understanding and predicting consequences in real physical environments, not just interpreting appearances the way a typical vision model does.

What are the different variants of NVIDIA Cosmos 3 and their uses?

Cosmos 3 Super (64B) is built for high-fidelity world modeling, Cosmos 3 Nano (16B) is designed for efficient reasoning and post-training, and Cosmos 3 Edge (4B) is meant for lightweight edge GPU deployment, including on robots running Jetson Thor hardware.

How is the NVIDIA Cosmos Coalition supporting physical AI development?

It fosters collaboration among world model builders and AI developers globally, and it has recently expanded to include robotics and manufacturing leaders in Japan, who plan to develop open world models for sectors ranging from logistics to healthcare.

Article produced with the assistance of artificial intelligence and reviewed by the editorial team.