Inside NVIDIA Cosmos 3: Physical Reasoning, World Models and Action Models

NVIDIA Cosmos 3 is not only a video generation model. It is presented as a foundation model for physical AI, combining physical reasoning, world generation and action generation in one open system.

This is important because robots and autonomous systems need more than visual perception. They must understand what is happening, predict what is likely to happen next and choose actions adapted to a specific environment.

According to NVIDIA’s technical blog, Cosmos 3 uses a Mixture-of-Transformers architecture with two main parts. The first is a Reasoner tower, which interprets multimodal inputs such as images, videos and text. The second is a Generator tower, which produces future observations, videos and action sequences.

This architecture matters because it connects understanding and generation. The model can reason about motion, object interactions and physical context before generating a prediction or an action-related output.

NVIDIA also released Cosmos 3 Nano and Cosmos 3 Super. Nano is designed for efficient inference, while Super targets higher-quality physical reasoning and generation for more demanding use cases.

Another important part of the release is the availability of open datasets for robotics, autonomous driving, warehouse operations, physical interaction, spatial reasoning and human motion. These datasets can help developers train and adapt world models for real-world physical AI applications.

For startups and researchers, the message is clear: physical AI will require better simulation, synthetic data and action-aware models.

Cosmos 3 shows that world models are becoming a serious infrastructure layer for robot learning, autonomous vehicles and embodied intelligence.

Source: NVIDIA Developer Blog, “Develop Physical AI Reasoning, World, and Action Models with NVIDIA Cosmos 3,” May 31, 2026.