Skip to content
RobothecaBeta

Cosmos 3

  • AI model
  • Humanoids and robot learning

An open family of omnimodal world foundation models from NVIDIA for physical AI, covering robots, autonomous vehicles and smart infrastructure.

  • Custom weights terms

Made by NVIDIA, United States (checked on github.com/NVIDIA (opens in a new tab)).

Source on GitHub: github.com/NVIDIA/Cosmos (opens in a new tab). Facts read at commit 8e2f3c7 (opens in a new tab).

Licenses

Code
OpenMDW 1.1Terms of their own, not a standard license

Read by Robotheca from LICENSE (opens in a new tab) at commit 8e2f3c7 (opens in a new tab), on 30 Oct 2026. A license text Robotheca's reader does not know: a person at Robotheca set it on 5 Nov 2026.

Weights
OpenMDW 1.1Terms of their own, not a standard license

Read on huggingface.co/nvidia/Cosmos3-Edge (opens in a new tab).

Commercial use
Yes

Code and weights are under OpenMDW-1.1, which allows commercial use. Running it needs NVIDIA's gated Cosmos-1.0-Guardrail model, under the NVIDIA Open Model License (commercial use allowed; revocable).

Labels

  • Custom weights termsThe weights come under their maker's own terms (OpenMDW 1.1), not a standard license: read their conditions before any use.

What it does: Robotheca's summary

  • Cosmos 3 is a Mixture-of-Transformers suite that supports two surfaces: a Reasoner for text and vision understanding, and a Generator for video, sound and action outputs.
  • The Generator covers world simulation, future prediction, synthetic data and robot policy learning.
  • Ships as three base models: Cosmos3-Super (64B), Cosmos3-Nano (16B) and Cosmos3-Edge (4B) for on-device use.
  • Example post-trained checkpoints demonstrate text-to-image, image-to-video and DROID robot policies.
  • Includes runnable notebooks, a Diffusers quickstart for video generation, and evaluation suites such as PAIBench and Physics-IQ.
  • Training and post-training are handled by the separate Cosmos Framework project, and data curation by Cosmos Curator.

Written by Robotheca from the project's README (opens in a new tab) at commit 6e2c69a.

Getting started

Difficulty
Advanced
Hardware it needs, as the project states it
Super runs on H200, B200 or GB200; Nano runs on RTX Pro 6000, H100 or B200; Edge runs on Jetson AGX Orin, Jetson Thor or RTX Pro 6000. Diffusion generation is described as compute-heavy.

Maintenance

Status
Active: a commit in the last 90 days
Newest commit
30 Oct 2026 on main
Latest release
Cosmos 3 Launch (1 Jun 2026)

Checked on GitHub on 30 Oct 2026.

Links

Robots that can run it

No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.

Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.