Cosmos 3: Omnimodal World Models for Physical AI

2026-06-01

AI Summary

This work proposes Cosmos 3, the first unified world model capable of end-to-end joint modeling of language, images, video, audio, and action sequences to support multimodal perception and decision-making for embodied agents in the physical world. Built upon a hybrid Transformer architecture, Cosmos 3 integrates visual-language understanding, video generation, world simulation, and action policy learning within a single framework. Scalability is achieved through large-scale synthetic data training and a unified input–output interface. The model achieves state-of-the-art performance across multiple multimodal understanding and generation benchmarks and has been recognized by Artificial Analysis as the top open-source text-to-image and image-to-video model, and by RoboArena as the leading policy model.

Abstract

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI -- effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 https://openmdw.ai/license/1-1/ License at https://github.com/nvidia/cosmos}{github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3 . The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3 .

Problem

• omnimodal
• world models
• Physical AI
• multimodal integration
• embodied agents

Innovation

• omnimodal
• world models
• mixture-of-transformers
• Physical AI
• embodied agents