NVIDIA Corporation
香港商輝達香港控股有限公司台灣分公司
NVIDIA Cosmos
Winning Reason
The Cosmos platform gives a boost to the development of physical AI, including robotics and autonomous driving, introducing open and customizable inference models, bringing together advanced generative world models, high-performance visual tokenizers, security protection mechanisms, and video processing pipelines to provide powerful tools for AI developers. By adopting the Nvidia Cosmos Tokenizer as the visual tokenizer, it can efficiently convert images and videos into computer-processable tokens for excellent performance boosts in both overall compression rate and processing speed.
Product Feature
Cosmos World Foundation Models are a family of open models trained on 9000 trillion tokens of robotics and autonomous driving data, providing high-quality data for physical AI training.
Available in 4B to 14B parameters, Cosmos supports autoregressive and diffusion models, giving developers flexibility based on their infrastructure and system needs.
Advanced controls let developers guide outputs using instruction videos to create domain-specific synthetic data. They can modify layout, motion, trajectory, and visual quality of output to produce diverse datasets that precisely enhance and accelerate AI training.
Built-in guardrails prevent hallucinations and harmful content. Deployed as an NVIDIA NIM, Cosmos enables fast, containerized AI deployment for robotics and autonomous systems.