NVIDIA Cosmos

What it does

Cosmos 3 is a family of open world models that take in text, images, video, ambient sound and action data. A reasoner tower reads text and vision and outputs text, covering scene understanding, grounding and task planning. A generator tower outputs video, audio and action, covering world simulation, future prediction, synthetic data and robot policy learning. Three base sizes (Super 64B, Nano 16B, Edge 4B) are published, alongside open datasets and supporting tools for curating data, evaluating outputs and training models.

Developers download checkpoints from Hugging Face and run them through Diffusers, Transformers, vLLM, SGLang, TensorRT-LLM or NIM, including an OpenAI-compatible serving option. Typical uses include generating or augmenting driving data with new weather and lighting, post-training a model on a specific robot's cameras and tasks, running closed-loop simulation, and building video-analytics agents that caption or search footage. The distinguishing point is that models, training framework, curation and evaluation tools are all open, with agent-runnable cookbook recipes.

It suits robotics, autonomous-vehicle and vision-AI developers comfortable with Python and GPUs. Models are released under OpenMDW-1.1. Running the quickstart requires an NVIDIA GPU, Hugging Face access to the gated Guardrail repository, and a large checkpoint download. Outputs can show temporal inconsistency or implausible physics, and safety-critical uses need extra validation.

What it can do

• Reason over video and images as a vision language model
• Generate video, audio and action from text, image, video, sound and action inputs
• Post-train models for robot policies and specialized camera setups
• Run closed-loop world simulation for driving and robot policies
• Curate and filter multimodal training data with Cosmos Curator
• Score generated synthetic video with Cosmos Evaluator

Works with

• Hugging Face
• Diffusers
• vLLM
• SGLang
• TensorRT-LLM
• NVIDIA NIM
• NVIDIA Isaac Lab
• NVIDIA Omniverse NuRec

Example workflows

Augment driving data

Input: existing driving video or scenario layouts. Steps: post-train or use a transfer checkpoint, generate variants with new weather, lighting or views, then score outputs with Cosmos Evaluator. Output: synthetic multi-view driving clips for training and testing.

Specialize a robot policy

Input: camera and embodiment data from a specific robot. Steps: start from a Cosmos 3 base model, post-train with the Cosmos Framework (SFT or LoRA), then test in closed-loop simulation. Output: a task-specific policy model.

Before you choose

• The first run downloads a 16B checkpoint, and diffusion steps are compute-heavy, so GPU hardware is needed
• Outputs can show artifacts or implausible dynamics; safety-critical use needs additional validation
• Guardrails are on by default and require gated repository access

Listed under

• Coding
• Sandboxes and infrastructure

Get started

Quickstart needs Guardrail repo access and a GPU

1. Request access to the gated nvidia/Cosmos-1.0-Guardrail repository on Hugging Face and accept its conditions
2. Create a Python environment, install the listed dependencies and log in with a Hugging Face read token
3. Run the sample Cosmos3-Nano script on a CUDA GPU to generate a first video

The source code is available under OpenMDW-1.1. GPU compute and deployment services are separate; review model license terms.

Source: github.com/nvidia/Cosmos, checked 28 Oct 2026. Prices can change; check before you buy.

Sources for this overview

Product pages, documentation, and pricing checked 28 Oct 2026. Each source supports part of the overview; availability and prices can change.

• NVIDIA Cosmos: Physical AI Development Platform ↗www.nvidia.com
• GitHub - NVIDIA/cosmos: NVIDIA Cosmos is an open platform of world models, datasets, and tools that enables developers to build Physical AI for robots, autonomous vehicles, smart infrastructure, and more. · GitHub ↗github.com

Company website

nvidia.com: product, pricing and policy pages

Web search

Confirms the official site

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

Last checked 28 Oct 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.