AI BriefingNVIDIA Cosmos 3
Key point
NVIDIA has open-sourced Cosmos 3, a frontier foundation model for Physical AI.
Details
NVIDIA has unveiled Cosmos 3, a frontier foundation model for Physical AI. This model integrates physical reasoning, world generation, and action generation within a single open model, designed to enable robots, autonomous vehicles, smart spaces, and more to understand, predict, and act in the real world.
Cosmos 3 is built on a Mixture-of-Transformers (MoT) architecture, and uses a Reasoner tower (VLM-based) as its core component to interpret multimodal observation data such as images, video, and text.
To enhance openness and reproducibility in Physical AI development, NVIDIA is providing the following resources as open source:
- Cosmos 3 Nano and Super model checkpoints (Hugging Face)
- Model training scripts and deployment tools (GitHub)
- Open datasets for robotics and autonomous driving
- Cosmos NIM microservices supporting optimized deployment on NVIDIA GPUs
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.