Robotics & Embodied AI

NVIDIA Cosmos 3

The physical-AI omnimodel that made license nerds sit up

Open code / restricted weightsModel

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.

Related & alternatives