NVIDIA Cosmos

NVIDIA Cosmos provides world models for developers building robots, autonomous vehicles and industrial vision systems. It supports reasoning about physical scenes, generating training data and learning actions before real-world deployment.

• Visual reasoning: Analyze objects, interactions and intent in images or video, with video search and summarization through NVIDIA Metropolis Blueprint.
• Synthetic video: Generate possible future scenes from text, images, video, ambient sound and actions to expand training data beyond recorded captures.
• Simulation and robot policies: Compare behaviors in controllable world simulations and adapt policy models to specific tasks, robot embodiments and camera layouts.
• Training and evaluation: Use Cosmos Framework for model training and optimization, Cosmos Curator for sensor-data filtering and annotation, and Cosmos Evaluator to score generated outputs.

Downloadable models run on your own NVIDIA hardware. Browser access uses NVIDIA's hosted catalog. Edge models support Jetson AGX Orin and Thor; larger models target workstation or data-center GPUs such as RTX PRO 6000, H100, H200 and B200. Backends include Diffusers, Transformers, vLLM, TensorRT-LLM and NIM, with OpenAI-compatible API serving available.

The video quickstart requires approved Hugging Face access to a guardrail model. Generation enables guardrails by default, but outputs can still show inconsistent motion, changing object shapes or implausible physics. Safety-critical uses require additional validation.