News
General
2 views
Nvidia unveils Cosmos 3 open foundation model for physical AI
Jun 02, 2026
📍 Philadelphia, PA, USA
🤖🌍 Nvidia has unveiled **Cosmos 3**, a groundbreaking open-world foundation model that could dramatically accelerate the development of robots, autonomous vehicles, and next-generation physical AI systems. Described by the company as the first fully open **“omnimodel,”** Cosmos 3 can natively understand and generate **text, images, video, ambient sound, and actions** while modeling how objects and machines behave in the real world with advanced physics-based accuracy.
According to Nvidia, Cosmos 3 was trained on an enormous dataset of **20 trillion multimodal tokens**, including nearly a billion images, hundreds of millions of videos, audio recordings, text, and human and robotic action data. Unlike traditional AI models that primarily generate content, Cosmos is designed to understand how machines move, interact, and make decisions in physical environments, making it especially valuable for robotics and autonomous systems.
Nvidia CEO **Jensen Huang** called the launch a major step toward the future of physical AI, saying breakthroughs in multimodal reasoning and world models are bringing intelligent robots and autonomous machines closer to reality. The company believes Cosmos can reduce AI training and testing cycles from months to days by allowing developers to simulate real-world scenarios digitally before deployment.
One of Cosmos 3’s most powerful capabilities is generating rare or dangerous situations—such as vehicle accidents, unusual road events, or robot collisions—that are difficult, expensive, or unsafe to capture in real life. This could significantly improve how AI systems learn to handle unpredictable environments while reducing development costs.
Nvidia is also launching the **Nvidia Cosmos Coalition**, bringing together AI developers and world-model researchers to advance the next generation of intelligent machines. With versions optimized for high-precision robotics, ultra-fast inference, and future edge devices, Cosmos 3 signals Nvidia’s ambition to become a foundational platform for the emerging era of physical AI. 🚀⚙️🌐
According to Nvidia, Cosmos 3 was trained on an enormous dataset of **20 trillion multimodal tokens**, including nearly a billion images, hundreds of millions of videos, audio recordings, text, and human and robotic action data. Unlike traditional AI models that primarily generate content, Cosmos is designed to understand how machines move, interact, and make decisions in physical environments, making it especially valuable for robotics and autonomous systems.
Nvidia CEO **Jensen Huang** called the launch a major step toward the future of physical AI, saying breakthroughs in multimodal reasoning and world models are bringing intelligent robots and autonomous machines closer to reality. The company believes Cosmos can reduce AI training and testing cycles from months to days by allowing developers to simulate real-world scenarios digitally before deployment.
One of Cosmos 3’s most powerful capabilities is generating rare or dangerous situations—such as vehicle accidents, unusual road events, or robot collisions—that are difficult, expensive, or unsafe to capture in real life. This could significantly improve how AI systems learn to handle unpredictable environments while reducing development costs.
Nvidia is also launching the **Nvidia Cosmos Coalition**, bringing together AI developers and world-model researchers to advance the next generation of intelligent machines. With versions optimized for high-precision robotics, ultra-fast inference, and future edge devices, Cosmos 3 signals Nvidia’s ambition to become a foundational platform for the emerging era of physical AI. 🚀⚙️🌐
Tags
news
Comments (0)
Login to post comments
No comments yet
Be the first to share your thoughts about this post.