At CES 2025 in Las Vegas, Nvidia introduced its new line of AI models called Cosmos World Foundation Models (Cosmos WFMs).
These models are designed to predict and generate “physics-aware” videos, making them ideal for a variety of applications like robotics, autonomous vehicles, and industrial AI.
The Cosmos WFMs come in three categories: Nano, Super, and Ultra. Each model is built to cater to different needs, from low-latency real-time applications to high-performance outputs.
The models range from 4 billion to 14 billion parameters, with larger models offering better performance.
These models are available for free use by developers and researchers through Nvidia’s API, GitHub, and Hugging Face.
Nvidia also offers fine-tuned models for specific tasks, such as generating sensor data for self-driving cars. Additionally, the company has released an upsampling model for augmented reality and guardrail models to ensure responsible usage.
Nvidia trained the Cosmos models on 9,000 trillion tokens from 20 million hours of real-world data, including human interactions, environmental data, and industrial use cases.
While the company claims the models are not designed to infringe on copyrighted works, some reports have raised concerns about the training data’s origins.
Despite this, Nvidia believes the Cosmos models will play a significant role in AI research and development, particularly in fields like robotics and autonomous driving.
Companies like Waabi, Wayve, Fortellix, and Uber are already piloting these models for various applications.
Nvidia CEO Jensen Huang expressed confidence that Cosmos will revolutionize robotics and industrial AI, similar to how Llama has impacted enterprise AI.
