NVIDIA Releases Cosmos 3 for World Simulation

NVIDIA has introduced Cosmos 3, the latest generation of its open models aimed at interactive world simulation and video generation. According to the company's announcement on Hugging Face, the models can generate video from both text prompts and input images, extending NVIDIA's ongoing effort to build foundation models that understand and predict physical environments.

What it does

Cosmos sits in a growing category of "world models" — systems designed to simulate how a scene evolves over time rather than simply producing a static image. That focus on interactivity and continuity makes the family especially relevant to robotics, autonomous systems, and other domains where machines need a sense of how the world behaves.

The release emphasizes two core capabilities:

• Image-to-video, turning a single frame into a coherent generated sequence
• Text-to-video, producing motion and scenes directly from written prompts

The models are published under a custom license, continuing NVIDIA's pattern of making weights broadly available while attaching its own usage terms rather than a standard open-source license.

Why it matters: High-quality simulated environments are a bottleneck for training physical AI. By releasing Cosmos 3 openly, NVIDIA gives researchers and developers a way to generate synthetic worlds and video data without building such systems from scratch — a meaningful step as interactive simulation becomes central to robotics and embodied AI. Full details are available in NVIDIA's Cosmos 3 announcement.