Step-Video-T2V

Training compute

4.1×10²⁴ FLOP

Parameters

30B

Published

Feb 24, 2025

Step-Video-T2V is an AI model developed by StepFun (China), first published in February 2025. It works in the video domain, on tasks such as video generation and text-to-video.

Training it took an estimated 4.1×10²⁴ FLOP of compute (estimation method: hardware). The model has 30,000,000,000 parameters. Training ran on 5,000 NVIDIA H800 SXM5 for about 720 hours.

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

Full record

Organization: StepFun
Country of organization: China
Domain: Video
Task: Video generation, Text-to-video
Training compute: 4.1×10²⁴ FLOP
Compute estimation method: Hardware
Parameters: 30,000,000,000
Training hardware: NVIDIA H800 SXM5
Chips used: 5,000
Training time: 720 h
Training power draw: 6.9 MW
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Epoch confidence: Likely