Title: Cosmos 3: Omnimodal World Models for Physical AI

Abstract:

We present Cosmos 3, a suite of omnimodal world models built on a unified mixture-of-transformers architecture that simultaneously handles the processing and generation of language, audio, images, video, and action sequences. This framework supports versatile input-output configurations, effectively consolidating vision-language models, video generation tools, world simulators, and world-action models into a single, cohesive system tailored for Physical AI.

Our assessment reveals that Cosmos 3 achieves new state-of-the-art performance across a wide range of understanding and generation tasks, proving its viability as a scalable, general-purpose backbone for embodied agents. At the time this technical report was finalized, our post-trained Cosmos 3 models received top honors from Artificial Analysis as the leading open-source solutions for both Text-to-Image and Image-to-Video generation, and were recognized by RoboArena as the premier policy model.

To foster open research and deployment in the field of Physical AI, we are releasing our code, model checkpoints, curated synthetic datasets, and evaluation benchmarks under the Linux Foundation's OpenMDW-1.1 license (https://openmdw.ai/license/1-1/). These resources are accessible at https://github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3. Further details can be found on the project website: https://research.nvidia.com/labs/cosmos-lab/cosmos3.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC