NVIDIA/cosmos

What is it?

Key features

• 01Reasoner model for visual and text understanding, grounding and embodied task planning
• 02Generator model for video, sound and robot action generation from multimodal inputs
• 03Three model sizes from 4B edge to 64B for different hardware tiers
• 04Runnable cookbook notebooks covering generation, transfer, action policy and reasoning
• 05Evaluation suites including PAIBench, Physics-IQ and VLMEvalKit
• 06Runs on Diffusers, Transformers, vLLM, SGLang, TensorRT-LLM and NIM

Use cases

• Generating synthetic driving or robot training video in simulated scenarios
• Running a robot manipulation policy on edge hardware such as Jetson
• Serving an OpenAI-compatible visual reasoning API for embodied applications
• Evaluating world model outputs against standard physics and quality benchmarks

Quick start

1. 1

   Create a Python 3.13 virtual environment and activate it

   uv venv --python 3.13 --seed --managed-python && source .venv/bin/activate
2. 2

   Install the required packages with uv

   uv pip install --torch-backend=auto "diffusers @ git+https://github.com/huggingface/diffusers.git" accelerate av cosmos_guardrail huggingface_hub imageio imageio-ffmpeg torch torchvision transformers
3. 3

   Log in to Hugging Face to access the gated Guardrail repository (request access first)

   uvx hf@latest auth login
4. 4

   Run the Python example from the README to generate your first video

Things to know

• !Access to the gated nvidia/Cosmos-1.0-Guardrail repository must be requested and approved before the code runs.
• !The Nano checkpoint is 16B parameters and needs a high-end GPU such as an RTX Pro 6000, H100 or B200.
• !Source code and models use the OpenMDW-1.1 license, with custom licensing available by contacting NVIDIA.

Topics

FAQ

What is cosmos?

Open world foundation models, datasets and tools from NVIDIA for building physical AI in robots, autonomous vehicles and smart infrastructure. NVIDIA Cosmos is an open platform for physical AI developers. Its current model family, Cosmos 3, is a set of omnimodal world models that come in two surfaces: a Reasoner that takes text and vision and outputs text for scene understanding and task planning, and a Generator that takes text, vision, sound and action and outputs video, sound or robot actions for simulation, future prediction and synthetic data.

Is cosmos free to use? What license does it have?

cosmos is open source under the Other license. Check the license terms before using it in commercial products.

Is cosmos still maintained?

Actively maintained. The last commit to cosmos was on Oct 16, 2026.

What can I use cosmos for?

Generating synthetic driving or robot training video in simulated scenarios Running a robot manipulation policy on edge hardware such as Jetson Serving an OpenAI-compatible visual reasoning API for embodied applications Evaluating world model outputs against standard physics and quality benchmarks

How do I install cosmos?

Create a Python 3.13 virtual environment and activate it: uv venv --python 3.13 --seed --managed-python && source .venv/bin/activate Install the required packages with uv: uv pip install --torch-backend=auto "diffusers @ git+https://github.com/huggingface/diffusers.git" accelerate av cosmos_guardrail huggingface_hub imageio imageio-ffmpeg torch torchvision transformers Log in to Hugging Face to access the gated Guardrail repository (request access first): uvx hf@latest auth login Run the Python example from the README to generate your first video

Where can I get cosmos?

The source code and installation instructions are on GitHub: https://github.com/NVIDIA/cosmos