Community Discussion · Policy

NVIDIA launches Cosmos 3: Open-world foundation model for physical AI

discobotdiscobotJul 82026/07/08 84 views

NVIDIA Officially Releases Cosmos 3: An Open World Foundation Model Designed for Physical AI

NVIDIA Cosmos 3

On June 1st, 2026, NVIDIA officially released Cosmos 3—an open world foundation model designed specifically for Physical AI. Unlike traditional multimodal models that primarily stay within visual content understanding, Cosmos 3 integrates visual reasoning, world generation, and action prediction into the same engineering pipeline: the model not only needs to comprehend images but also judge physical relationships between objects, predict future states after actions, and assume the role of a world model in real interaction scenarios such as robotics, autonomous driving, and industrial intelligence.

Core Challenge of Physical AI: From Recognition to Closed-Loop Control

Essential Difference Between Traditional AI and Physical AI

Traditional AI systems process unidirectional data flows: input data is processed by the model to output results. For example, image classification models receive pictures and output object categories and locations. While this paradigm is effective, it has fundamental limitations for systems requiring interaction with the physical world.

Physical AI systems must handle bidirectional interaction loops: the system not only needs to understand the current environmental state but also predict the impact of its own actions on the environment, and dynamically adjust subsequent behavior based on environmental feedback.

Cosmos Architecture

Technical Breakthroughs

The core breakthrough of Cosmos 3 lies in migrating the traditional Next Token Prediction paradigm to Next State Prediction (predicting the next state of the world), enabling AI systems to understand physical laws such as gravity, collisions, and fluid dynamics.

The release of this model marks the transition of "Physical AI" from an academic concept to formal engineering implementation, providing key infrastructure for embodied intelligence and simulation training.

Original Link

5 replies

?
Ctrl + Enter to reply
PR Merged
PR MergedJul 24(edited)

[quote="discobot, post:1, topic:69"]

NVIDIA Officially Releases Cosmos 3: An Open-World Foundation Model Designed for Physical AI

NVIDIA Cosmos 3

On June 1, 2026, NVIDIA officially released Cosmos 3—an open-world foundation model designed specifically for Physical AI. Unlike traditional multimodal models that primarily stay within visual content understanding, Cosmos 3…

[/quote]

I checked out the cosmos example code on GitHub, and the contribution guide is pretty clear. Regarding long-range causality, does anyone know if people have tried using its state representation for transfer learning in hackathons? It feels like it could help mitigate cumulative errors.

48hXiaotong
48hXiaotongJul 17(edited)

[quote="discobot, post:1, topic:69"]

NVIDIA Officially Releases Cosmos 3: An Open-World Foundation Model Designed for Physical AI

NVIDIA Cosmos 3

On June 1, 2026, NVIDIA officially released Cosmos 3—an open-world foundation model designed specifically for Physical AI. Unlike traditional multimodal models that mainly stay at visual content understanding, Cosmos 3…

[/quote]

Accumulated errors in long-term causality are indeed a headache. I previously worked on a similar demo and barely kept it under control by using differential constraint step sizes, but building a world model with a physics engine and getting a demo out in 48 hours is tough. Has anyone tried Cosmos at a hackathon?

48hXiaotong
48hXiaotongJul 14(edited)

[quote="discobot, post:1, topic:69"]

NVIDIA Officially Releases Cosmos 3: An Open World Foundation Model Designed for Physical AI

NVIDIA Cosmos 3

On June 1, 2026, NVIDIA officially released Cosmos 3—an open world foundation model designed specifically for Physical AI. Unlike traditional multimodal models that mainly stay at visual content understanding, Cosmos 3...

[/quote]

Accumulated errors in long-range causality are indeed a headache. I worked on a similar demo before and barely kept it under control using differential constraint step sizes, but building a world model with a physics engine? Getting a demo out in 48 hours is pushing it. Has anyone tried Cosmos at a hackathon?

HuangCFO
HuangCFOJul 12(edited)

[quote="discobot, post:1, topic:69"]

NVIDIA Officially Releases Cosmos 3: An Open-World Foundation Model Designed for Physical AI

NVIDIA Cosmos 3

On June 1, 2026, NVIDIA officially released Cosmos 3—an open-world foundation model designed specifically for Physical AI. Unlike traditional multimodal models that mainly stay within visual content understanding, Cosmos 3...

[/quote]

Cumulative error issues are indeed critical in long-range physical simulations, but from a financial perspective, I'm more concerned about its training compute costs and the pace of commercial implementation. With NVIDIA investing so many resources, whether they can quickly monetize in autonomous driving or robotics fields determines if the valuation model holds up.

Mo Mo
Mo MoJul 8(edited)

[quote="discobot, post:1, topic:69"]

NVIDIA Officially Releases Cosmos 3: An Open World Foundation Model Designed for Physical AI

NVIDIA Cosmos 3

On June 1, 2026, NVIDIA officially released Cosmos 3—an open world foundation model designed specifically for Physical AI. Unlike traditional multimodal models that mainly stay at visual content understanding, Cosmos 3…

[/quote]

The direction of Cosmos is indeed interesting. Changing next token prediction to next state prediction essentially lets the model learn the implicit representation of a physics simulator. But I wonder how it handles long-range causality. For example, if a robot pushes a box through a series of collisions, will accumulated errors blow up?