NVIDIA Cosmos 3
The physical-AI omnimodel that made license nerds sit up
What it is
A unified world model for physical AI: video generation, physical reasoning, and action generation in one architecture. Ships as Nano (8B reasoner + 8B generator) for efficient inference and Super (32B + 32B) for large-scale synthetic data, with a cookbook, training framework, and post-training scripts across the Cosmos ecosystem.
Why it's interesting
Top of the robotics benchmarks, and the license is the other story: NVIDIA moved Cosmos to the Linux Foundation's OpenMDW-1.1 — a genuinely permissive model license granting rights 'without restriction' — rather than its own restriction-laden Open Model License.
Use cases
- Synthetic data generation for robot and AV training
- Physical reasoning over video for embodied agents
- Post-training world models for specific robots and environments
Who it's for
Robotics teams, AV researchers, simulation and synthetic-data engineers
Setup
Advanced. Nano wants a serious GPU; Super is datacenter territory. The ecosystem spans several repos — start from the Cosmos cookbook
Limitations & cautions
Datacenter-class hardware for the Super variants, an ecosystem sprawling across a dozen repos with mixed licenses and generations, and OpenMDW itself is young — permissive on paper, untested in court.
Editorial takeaway
When the biggest chipmaker on earth picks a brand-new license for its crown-jewel models, the license is the news. This one reads better than almost any corporate model license before it.