BLIP-2 (Q-Former)

Training compute

1.2×10²¹ FLOP

Parameters

1.5B

Published

Jan 30, 2023

BLIP-2 (Q-Former) is an AI model developed by Salesforce Research (United States), first published in January 2023. It works in the vision and language domain, on tasks such as visual question answering and image captioning.

Training it took an estimated 1.2×10²¹ FLOP of compute (estimation method: hardware). The model has 1,480,000,000 parameters. It was trained on roughly 2.3B datapoints. Training ran on 16 NVIDIA A100 SXM4 40 GB for about 200 hours. The compute alone is estimated at $2K in 2023 dollars.

Access: Open weights (unrestricted). Its weights are openly available. The reference paper has 8,019 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Salesforce Research
Country of organization: United States
Domain: Vision, Language
Task: Visual question answering, Image captioning
Training compute: 1.2×10²¹ FLOP
Compute estimation method: Hardware
Parameters: 1,480,000,000
Dataset size: 2.3B
Training hardware: NVIDIA A100 SXM4 40 GB
Chips used: 16
Training time: 200 h
Chip-hours: 3.2K
Training power draw: 12.8 kW
Training cost (2023 USD): $2K
Numerical format: FP16
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Citations: 8,019
Epoch confidence: Confident