Breaking

NVIDIA Releases Cosmos 3, an Open World Model for Physical AI

June 11, 2026 at 12:04 EDT

Foundation Models
Robotics
Open Source

NVIDIA on May 31, 2026 announced and began shipping NVIDIA Cosmos 3, an omnimodal open frontier world foundation model for physical AI such as robots and autonomous vehicles. CEO Jensen Huang introduced it at COMPUTEX 2026, with model weights released openly the same day. According to the official announcement, Cosmos 3 was developed by Cosmos Labs.

May 31, 2026 · NVIDIA · COMPUTEX / GTC Taipei

Cosmos 3: The Open Frontier Foundation Model for Physical AI

A "fully open omnimodel" that reasons, generates language, images, video, audio — and actions — in a single model, built to be the backbone for robots and autonomous vehicles. Released open-weight the same day.

single model unifying reasoning + world + action generation

leaderboard top spots among open models

synthetic data generation datasets released

model sizes: data center → workstation → edge

One family, scaled by parameters

Each block ≈ 8B parameters. Same model, tuned from data center down to the edge.

64B

Super

Data center

16B

Nano

Workstation

Edge

Coming soon

Closing the Physical AI loop

SYNTHETIC DATA

generate at scale

→

SIMULATION

world generation

→

POLICY TRAINING

robot actions

Targets the biggest bottleneck in Physical AI — the cost and time of collecting real-world data.

Nano runs on desktop-class hardware

7.8 GB

VRAM used

5.2 s

per frame

450

frames (15s clip)

Tracking succeeded on a real (not synthetic) highway clip on ARM Blackwell desktop systems.

Why it matters

Open weights under commercial OpenMDW 1.1 license
One model replaces separate reasoning + world + policy models
Open answer to closed models from OpenAI, Google, Amazon
Called a "ChatGPT moment for open Physical AI"

Caveats

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.