---
title: "NVIDIA: Cosmos World Foundation Model Platform for Physical AI"
slug: "nvidia-cosmos-world-foundation-model-platform-for-physical-ai"
author: "Jeremy Weaver"
date: "2025-01-08 20:49:45"
category: "Premium"
topics: "Cosmos World Foundation Model Platform, Pre-training and Post-training Paradigms, Diffusion and Autoregressive Models, Physical AI Applications, Video Tokenizer and Safety Guardrails"
summary: "NVIDIA's Cosmos World Foundation Model platform for Physical AI uses a dual-stage training approach with diffusion and autoregressive models on a massive curated video dataset to create versatile foundation models that are fine-tuned for robotic manipulation, autonomous driving, and other tasks, featuring a novel video tokenizer and integrated safety measures."
banner: ""
thumbnail: ""
---

NVIDIA: Cosmos World Foundation Model Platform for Physical AI<br><br><iframe title="NVIDIA: Cosmos World Foundation Model Platform for Physical AI" allowtransparency="true" height="300" width="100%" style="border: none; min-width: min(100%, 430px);height:300px;" scrolling="no" data-name="pb-iframe-player" src="https://www.podbean.com/player-v2/?from=embed&i=5jjnc-17a0a17-pb&square=1&share=1&download=1&fonts=Arial&skin=1&font-color=auto&rtl=0&logo_link=episode_page&btn-skin=7&size=300" loading="lazy" allowfullscreen=""></iframe><br><br><p>Summary of <a href='https://d1qx31qr3h6wln.cloudfront.net/publications/NVIDIA%20Cosmos_3.pdf'>Read Full Report (PDF)</a></p>
<p>Introduces NVIDIA's Cosmos World Foundation Model (WFM) platform for Physical AI. Cosmos uses a pre-training and post-training paradigm, employing both diffusion and autoregressive models trained on a massive, curated video dataset (20M hours) to create generalist WFMs.</p>
<p>These are then fine-tuned for specialized Physical AI tasks like robotic manipulation and autonomous driving. The platform includes a novel video tokenizer for efficient processing and a guardrail system for safety.</p>
<p>Results demonstrate state-of-the-art performance across various benchmarks and applications.</p>