NVIDIA Cosmos3-Edge: 4B Open Omnimodal World Model for Video

NVIDIA Cosmos3-Edge is the newest member of the Cosmos 3 family of open omnimodal world models. The 4B Mixture-of-Transformers (MoT) model takes text, images, video and action trajectories as input and generates coherent text, images, video and action outputs, covering world simulation, future prediction and action reasoning for Physical AI applications.

What Is Cosmos 3

Cosmos 3 unifies understanding, generation, simulation and action in a single transformer: an autoregressive tower handles discrete text tokens while a diffusion transformer synthesizes images, video, audio and actions through iterative denoising. There is no separate text encoder — language and diffusion tokens run through one shared transformer, so the model reads its own architecture from the checkpoint.

Cosmos3-Edge supports text and image inputs at 256p and 480p with 16:9, 4:3, 1:1, 3:4 and 9:16 aspect ratios, and its action-conditioning works across robot embodiments such as Franka Panda arms, autonomous vehicles, drones and camera trajectories. The outputs below show action-conditioned audio-video generation from the model card:

ComfyUI Support

Cosmos3 is not part of ComfyUI core, but the community node pack ComfyUI-Cosmos3 supports the full Cosmos3 family — including Cosmos3-Edge for text-to-video and image-to-video — through the built-in KSampler / SamplerCustomAdvanced stack:

1. Clone ComfyUI-Cosmos3 into ComfyUI/custom_nodes/ and install its requirements.txt
2. Download a checkpoint (e.g. nvidia/Cosmos3-Edge) into ComfyUI/models/cosmos3/<name>/; the loader reads the architecture from transformer/config.json, so no manual setup is needed
3. Sample with the uni_pc_bh2 flow-matching schedule; the text tower is prefilled once per prompt and its K/V reused across denoising steps

The pack ships example workflows for text-to-video, image-to-video, joint audio-video and the 4-step distilled Super checkpoints:

Download the Cosmos3-Edge text-to-video workflow Download the Cosmos3 image-to-video workflow

Availability

Weights are open on Hugging Face, with the model collection at huggingface.co/collections/nvidia/cosmos3. Official inference and training code lives in the NVIDIA Cosmos repository and the Cosmos Framework, and the technical report is on the Cosmos 3 research page.