Overview
- NVIDIA has published Cosmos 3 Super (32B) and Cosmos 3 Nano (8B) on Hugging Face with model cards, downloads and Diffusers integration for generation pipelines.
- The model uses a Mixture-of-Transformers design that pairs a reasoning transformer with a generation transformer to process text, images, video, ambient sound and action outputs in one backbone.
- NVIDIA released synthetic-data generation datasets, post-training scripts on GitHub and NIM microservices to let developers create training data and deploy models on cloud and workstation GPUs.
- The company announced the NVIDIA Cosmos Coalition and named early partners such as Agile Robots, Runway and Skild AI to coordinate open research and industry adoption for robotics and autonomous vehicles.
- NVIDIA says Cosmos 3 leads several physical-AI leaderboards but those rankings are vendor-reported and awaiting independent verification, and the firm says a Cosmos 3 Edge build for real-time inference is coming soon.