NVIDIA · 2026-05-31 · major
NVIDIA Cosmos 3 — Open Mixture-of-Transformers Omni-Model for Physical AI Lands at Computex With Nano (16B) and Super (65B) Variants, OpenMDW 1.1 Weights, and a New Coalition With Runway, Black Forest Labs, and Skild AI
NVIDIA open-sourced Cosmos 3 at Computex Taipei, a Mixture-of-Transformers omni-model that unifies text, image, video, audio, and action reasoning in a single foundation model for physical AI, shipping under OpenMDW 1.1.
An open frontier omni-model that can reason about, simulate, and act in the physical world.
Key specs
| License | OpenMDW 1.1 |
|---|---|
| GitHub stars | 8.3k |
| Nano params | 16B (8B reasoner + 8B generator) |
| Super params | 65B (32B reasoner + 32B generator) |
What is it?
Cosmos 3 is NVIDIA's third-generation open world foundation model for physical AI. Unlike text-only LLMs, it natively understands and generates across five modalities — text, images, video, ambient audio, and robot actions — letting a single model reason about a scene, continue it as video, and emit a control policy. Two variants are open today, with Edge promised for later.
How does it work?
The architecture is a Mixture-of-Transformers that combines autoregressive and diffusion subsequences with separate parameter sets and joint attention, so one forward pass can mix discrete reasoning with continuous video and audio generation. Cosmos 3 Nano pairs an 8B reasoner with an 8B generator for inference on a workstation-class RTX PRO 6000, while Cosmos 3 Super pairs a 32B reasoner with a 32B generator for large-scale synthetic data generation on Hopper and Blackwell. NVIDIA released weights, training scripts, datasets, benchmarks, and code together under the OpenMDW 1.1 license.
Why does it matter?
Open frontier physical-AI models are still rare and Cosmos 3 ships a far more complete release than most labs offer — weights plus the data and training recipes. NVIDIA also launched a Cosmos Coalition with Runway, Black Forest Labs, Skild AI, Agile Robots, Generalist, and LTX to align on open world-model interfaces, signaling Cosmos as a likely default substrate for robotics and autonomous-vehicle teams that don't want to train a frontier omni-model in-house.
Who is it for?
robotics, autonomous-vehicle, and simulation teams
Try it
huggingface.co/nvidia/Cosmos3-Nano