This page may contain stale information. Last updated: 2026-06-01
Overview
NVIDIA Cosmos 3 is an open frontier foundation model for physical-ai announced at GTC Taipei on May 31, 2026. It is the first fully open “omnimodel” unifying vision reasoning, world generation, and action prediction in a single Mixture-of-Transformers (MoT) architecture.
Model Variants
| Model | Parameters | Target Hardware | Use Case |
|---|---|---|---|
| Cosmos 3 Nano | 16B | RTX PRO 6000 | Real-time robotics inference |
| Cosmos 3 Super | 64B | Hopper/Blackwell | Synthetic data generation, advanced reasoning |
| Cosmos 3 Edge | TBD | Edge devices | Real-time inference (coming) |
Architecture
- Reasoner tower: Autoregressive VLM interpreting multimodal inputs (text, images, video)
- Generator tower: Diffusion-based physics-aware video and action generation
Benchmarks
Leads open-source models on VANTAGE-Bench, PAI-Bench, R-Bench Physics-IQ, RoboLab, and Artificial Analysis text-to-image/image-to-video leaderboards. NVIDIA HUE (Human Evaluation) framework for fine-grained quality assessment.
Ecosystem
- Open weights on Hugging Face, code on GitHub
- Six SDG datasets (robotics, driving, warehouse, physics, human motion, spatial reasoning)
- Cosmos NIM microservices for production deployment
- Cosmos Coalition: Agile Robots, Black Forest Labs, Generalist, LTX, Runway, Skild AI
"Months to days" training cycle claims are NVIDIA-stated marketing — not independently verified.