Breaking

NVIDIA Releases Open Cosmos 3 Omnimodel for Physical AI; 64B Super Lands on Design Arena

June 28, 2026 at 19:00 EDT

Foundation Models
Open Source
Robotics

NVIDIA has released NVIDIA Cosmos 3, an open-weight foundation model aimed at AI that operates in the physical world, and its larger 64-billion-parameter variant, Cosmos 3 Super, is now available on the generation and benchmarking platform Design Arena. The model is an omnimodel that natively understands and generates across text, image, video, and audio, which NVIDIA describes as an "open frontier foundation model for physical AI."

May 31, 2026 · NVIDIA

Cosmos 3: One Open "Omnimodel" for Physical AI

NVIDIA's open-weight World Foundation Model takes text, image, video and action as inputs — and outputs text, image, video, audio and action — replacing separate vision, video, simulation and robot-policy models with a single architecture.

64B

Parameters in Cosmos 3 Super (open weights)

Modalities generated: text · image · video · audio · action

~130GB

Est. VRAM for Super — beyond consumer GPUs

Two models, split by use case

Column height = parameter count. Super for the data center, Nano for workstation and edge.

64B

Super

Data center (Hopper–Blackwell)

16B

Nano

Workstation / edge

One architecture: reason → generate → act

Autoregressive
Reasoner

→

Diffusion
Generator

→

Action
Output

Mixture-of-Transformers (MoT) unifies what were separate VLMs, video generators, world simulators and robot policies.

Where it leads

Tops several Physical AI leaderboards among open-weight models — robotics, smart space and driving averages, plus world generation and vision understanding. Image-to-Video quality well received.

The caveats

On the general Text-to-Image Arena it sits ~8th–11th among open models. Super's ~130GB VRAM pushes many toward Nano; pipeline integration — not video quality — is the early stumbling block.

Continue reading

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