Gemma 3 4B

Training compute

9.6×10²² FLOP

Parameters

4B

Published

Mar 12, 2025

Gemma 3 4B is an AI model developed by Google DeepMind (United States), first published in March 2025. It works in the language, vision and multimodal domain, on tasks such as language modeling/generation, question answering, translation and 4 more.

Training it took an estimated 9.6×10²² FLOP of compute (estimation method: operation counting). The model has 4,000,000,000 parameters. It was trained on roughly 4T datapoints. Training ran on 2,048 Google TPU v5e.

Access: Open weights (restricted use). Its weights are openly available. It is built on top of SigLIP 400M. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Google DeepMind
Country of organization: United States
Domain: Language, Vision, Multimodal
Task: Language modeling/generation, Question answering, Translation, Chat, Quantitative reasoning, Visual question answering, Code generation
Training compute: 9.6×10²² FLOP
Compute estimation method: Operation counting
Parameters: 4,000,000,000
Dataset size: 4T
Training hardware: Google TPU v5e
Chips used: 2,048
Training power draw: 904.3 kW
Model accessibility: Open weights (restricted use)
Open weights: Yes
Base model: SigLIP 400M
Epoch confidence: Confident