Cosmos-1.0- Diffusion-14B Video2World

Training compute

2.8×10²⁴ FLOP

Parameters

14B

Published

Jan 7, 2025

Cosmos-1.0- Diffusion-14B Video2World is an AI model developed by NVIDIA (United States), first published in January 2025. It works in the robotics, vision and video domain, on tasks such as robotic manipulation, self-driving car and video generation.

Training it took an estimated 2.8×10²⁴ FLOP of compute (estimation method: hardware). The model has 14,000,000,000 parameters. It was trained on roughly 9000T datapoints. Training ran on 10,000 NVIDIA H100 SXM5 80GB.

Access: Open weights (restricted use). Its weights are openly available. Epoch AI rates the confidence of this record as likely.

Full record

Organization: NVIDIA
Country of organization: United States
Domain: Robotics, Vision, Video
Task: Robotic manipulation, Self-driving car, Video generation
Training compute: 2.8×10²⁴ FLOP
Compute estimation method: Hardware
Parameters: 14,000,000,000
Dataset size: 9000T
Training hardware: NVIDIA H100 SXM5 80GB
Chips used: 10,000
Training power draw: 13.8 MW
Model accessibility: Open weights (restricted use)
Open weights: Yes
Epoch confidence: Likely