NVIDIA Launches Cosmos 3 — Open World Foundation Model for Physical AI

What Is Cosmos 3?

Previous NVIDIA Cosmos models were purpose-built: Cosmos Predict for future prediction, Cosmos Transfer for domain transfer, Cosmos Reason for physical AI understanding, and Cosmos Policy for action generation. Cosmos 3 unifies all of these capabilities into a single open omnimodel.

The model accepts text, images, video, ambient sound, and actions as input — and can generate any combination of those modalities as output. A single model checkpoint now covers the full pipeline: visual reasoning → physically accurate synthetic video generation → action output.

NVIDIA's core pitch: physical AI systems should think before they act. Cosmos 3 is the infrastructure that makes that possible at scale, reducing training and evaluation cycles from months to days.

Two Models: Nano and Super

Cosmos 3 ships in two configurations:

Nano is designed to run on workstation-class GPUs for real-time robotics applications. Super targets datacenter GPUs for high-fidelity synthetic data generation and the most demanding physical reasoning tasks.

Mixture-of-Transformers Architecture

The technical breakthrough in Cosmos 3 is its Mixture-of-Transformers (MoT) architecture. This design subsumes multiple traditionally separate architectures — vision-language models (VLM), world models, world action models (WAM), and vision-language-action (VLA) models — into a single unified system.

This means that post-training Cosmos 3 on a specific domain transforms it into a world action model capable of perceiving its environment, reasoning about physics, generating plausible futures, and producing appropriate actions — all from a single checkpoint.

NVIDIA Cosmos Coalition

Alongside the model launch, NVIDIA announced the Cosmos Coalition — a collaboration between world model builders and AI developers. Founding members include Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI. The coalition is focused on advancing the next generation of open world models for physical AI applications.

Open-Source Datasets

NVIDIA simultaneously released six synthetic data generation (SDG) datasets covering:

• Robotics manipulation
• Physics simulation
• Spatial reasoning
• Human motion
• Autonomous driving
• Warehouse environments

These datasets are designed for post-training Cosmos 3 and other physical AI models, and are available on Hugging Face alongside the model weights.

Why World Models Matter for Physical AI

Training robots and autonomous vehicles requires vast amounts of real-world data — which is expensive to collect and nearly impossible to gather for rare edge cases. Cosmos 3 solves this by generating physically accurate synthetic video and action scenarios at scale, letting teams train and evaluate models against conditions they couldn't easily reproduce in the real world.