Starstruk Logo MarkStarstruk

Cosmos3

NVIDIA/cosmosCosmos3Jun 1, 2026by mingyuliutw

AI Summary

NVIDIA has released Cosmos 3, a next-generation open family of omnimodal world foundation models designed for Physical AI. This release unifies multiple data modalities and introduces a new 'Reasoner + Generator' architecture to enable advanced simulation, robotics, and agent capabilities. It features multiple model sizes, including the compact Cosmos3-Nano and the high-capacity Cosmos3-Super, alongside specific models for policy learning and video generation.

Key Highlights

  • Unified omnimodal architecture supporting text, images, video, audio, and actions within a single model.
  • Integrated 'Reasoner + Generator' architecture for improved world understanding and generation.
  • Native support for robotics through Cosmos3-Policy models, including the DROID-based Nano-Policy.
  • Open release of models, code, datasets, evaluation benchmarks, and inference tooling for the community.
  • State-of-the-art performance across world understanding, generation, and robotics benchmarks.

New Features

  • Integrated Reasoner + Generator architecture for world understanding and generation.
  • Native support for robot action generation and policy learning via Cosmos3-Policy models.
  • Flexible generation capabilities including Text-to-Image, Image-to-Video, Video-to-World, and multimodal simulation.
  • Open release of evaluation benchmarks and inference tooling for the Physical AI community.
  • State-of-the-art open model performance across world understanding and robotics benchmarks.

Full Release Notes

# Cosmos 3 is Here 🚀

Today, we're excited to release Cosmos 3 — NVIDIA's next-generation family of open omnimodal world foundation models for Physical AI.

Cosmos 3 unifies language, images, video, audio, and actions within a single architecture, enabling developers to build agents that can understand, reason about, simulate, and act in the physical world. From world generation and simulation to robotics and embodied AI, Cosmos 3 serves as a general-purpose foundation model for Physical AI.

What's new:

- 🌍 Unified omnimodal world model supporting text, image, video, audio, and action modalities
- 🧠 Integrated Reasoner + Generator architecture for world understanding and world generation
- 🎬 Flexible generation across Text-to-Image, Image-to-Video, Video-to-World, and multimodal simulation tasks
- 🤖 Native support for robot action generation and policy learning through Cosmos3-Policy models
- 📈 State-of-the-art open model performance across world understanding, generation, and robotics benchmarks
- 🔓 Open release of models, code, datasets, evaluation benchmarks, and inference tooling for the Physical AI community

:book: [Read the Paper](https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf) | :point_right: [Download the Models](https://huggingface.co/collections/nvidia/cosmos3) | 🧑‍🍳 [Explore the Cosmos Cookbook](https://github.com/NVIDIA/cosmos/tree/main/cookbooks/cosmos3)

The Cosmos 3 release includes:

- Cosmos3-Nano (16B) – Compact omnimodal world foundation model optimized for efficient deployment and development.
- Cosmos3-Super (64B) – High-capacity world model for advanced reasoning, generation, simulation, and Physical AI applications.
- Cosmos3-Super-Text2Image – State-of-the-art text-to-image generation model built on Cosmos 3.
- Cosmos3-Super-Image2Video – High-fidelity image-to-video generation model with strong temporal consistency and controllability.
- Cosmos3-Nano-Policy-DROID – Open robot foundation model for learning manipulation and control policies directly from demonstrations.

Cosmos 3 represents a major step toward general-purpose world models that can perceive, reason, simulate, and act—bringing us closer to a future where Physical AI can learn from both the real world and generated worlds at scale.