Back to Search
C

Cosmos 3 Edge

ModelActive

Cosmos 3 Edge is a 4-billion parameter compact world model from NVIDIA, released on July 15, 2026 as a new lightweight addition to the Cosmos 3 family of omnimodal world foundation models. It is designed for on-device (edge) vision reasoning, enabling real-time understanding of surroundings, physical reasoning, and generation of robot actions and policies locally without relying on cloud inference. Built on NVIDIA Nemotron using a Mixture-of-Transformers (MoT) design, Cosmos 3 Edge combines an autoregressive reasoner with a diffusion-based generator for omnimodal processing across text, image, video, audio, and action modalities. The model uses a dense approximately 2-billion parameter transformer design (28 layers, 2048 hidden size, 16 attention heads) that was trained from scratch. Cosmos 3 Edge supports embodied systems that 'see, reason in real time, and predict robot actions'. It can be post-trained and adapted in approximately one day via the open Cosmos framework for specific robots and applications. The model was announced during NVIDIA CEO Jensen Huang's visit to Japan, alongside the expansion of the Cosmos Coalition with over 20 Japanese companies including FANUC, Yaskawa, Kawasaki Heavy Industries, Fujitsu, SoftBank, and Sony.

Details

Updated:7/17/2026

No structured details available.

Tags

world-modelphysical-aiedge-AINVIDIAvision-reasoningNemotronMoTrobotics

Relationships

Sources

https://www.nvidia.com/en-us/ai/cosmos/
website
Visit
https://nvidianews.nvidia.com/news/japans-robotics-and-manufacturing-leaders-build-on-nvidia-cosmos-to-advance-physical-ai-frontier
website
Visit
https://arxiv.org/abs/2606.02800
paper
Visit

Appears In

No related landscapes.