Galactica

Training compute

3.2×10²³ FLOP

Parameters

120B

Published

Nov 16, 2022

Galactica is an AI model developed by Meta AI (United States), first published in November 2022. It works in the language and biology domain, on tasks such as language modeling, question answering, mathematical reasoning and 2 more.

Training it took an estimated 3.2×10²³ FLOP of compute (estimation method: operation counting). The model has 120,000,000,000 parameters. It was trained on roughly 106B datapoints. Training ran on 128 NVIDIA A100 SXM4 80 GB. The compute alone is estimated at $591K in 2023 dollars.

Access: Open weights (non-commercial). Its weights are openly available. The reference paper has 1,017 citations. Epoch AI rates the confidence of this record as likely.

Full record

Organization: Meta AI
Country of organization: United States
Domain: Language, Biology
Task: Language modeling, Question answering, Mathematical reasoning, Medical diagnosis, Language modeling/generation
Training compute: 3.2×10²³ FLOP
Compute estimation method: Operation counting
Parameters: 120,000,000,000
Dataset size: 106B
Training hardware: NVIDIA A100 SXM4 80 GB
Chips used: 128
Training power draw: 102.4 kW
Training cost (2023 USD): $591K
Model accessibility: Open weights (non-commercial)
Open weights: Yes
Citations: 1,017
Epoch confidence: Likely