NVIDIA launched Cosmos 3 on June 1, 2026, an open world foundation model for physical AI built on a mixture-of-transformers architecture.
According to the NVIDIA Newsroom, the model targets physical AI applications. These include robots, autonomous vehicles, and smart spaces that must understand the real world before acting within it.
What Cosmos 3 Is
Cosmos 3 is a world foundation model. Not a language model.
It generates synthetic world data, the kind that helps machines reason about physical environments.
Its mixture-of-transformers architecture breaks from the standard dense transformer stack. Instead of pushing every input through the full network, it routes each one to specialized sub-models. The result is less compute per forward pass.
NVIDIA calls Cosmos 3 open, and that word carries weight here.
Developers can reach the model weights directly. That's what sets it apart from the closed-weight alternatives crowding the physical AI space.
Also Read: Sui Foundation Blames Upgrade Bugs for Three Costly Outages
Background
NVIDIA entered the physical AI model space in 2024 with its first Cosmos release. That version established a framework for training robots and autonomous systems on generated world data.
The Cosmos series competes with world model research from Google DeepMind and Meta's JEPA architecture program. A May 2026 benchmark found current frontier models brittle when tested on physical-world reasoning tasks.
NVIDIA's data center business has posted consecutive record quarters through early 2026. Cosmos 3 extends that infrastructure push into the model layer.
Also Read: ETH Loses Its Last Floor And Stares Down A Drop Toward $1,800
Companion Open-Source Release
Alongside Cosmos 3, NVIDIA released a collection of open-source agent tools and skills for physical AI, per a separate announcement.
Those tools cover perception, navigation, and manipulation tasks. They are designed to run on top of Cosmos 3 as an inference backbone.
NVIDIA also released technical documentation on the DSX OS platform, which it describes as open and modular software for operating AI factories at scale.
The combined release positions NVIDIA to own multiple layers of the physical AI stack, from model weights to agent tooling to operating software.
Read Next: OpenAI Hands Japan's Megabanks Its GPT-5.5 Cyber Defense Model To Withstand AI Hacks