ViT-G/14

Closed weights Google Brain,Google Research 1.8B parameters June 2021

No estimate

No hardware requirements for this model

The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.

On record

Full specification

Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.

Origin

Who built this model, where, and when it was published.

Organisation
Google Brain,Google Research
Organisation type
Industry,Industry
Country
United States of America
Published
8 June 2021
Authors
X Zhai, A Kolesnikov, N Houlsby, L Beyer

What it does

The problem areas the model was built for. A model can carry several of each.

Domain
Vision
Task
Image classification
Approach
Self-supervised learning
Numerical format
BF16

Size

How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.

Parameters
1.8B

Table 2 of paper

Training data
3,000,000,000 tokens

"For this study, we use the proprietary JFT-3B dataset, a larger version of the JFT-300M dataset used in many previous works on large-scale computer vision models [31, 18, 11]. This dataset consists of nearly 3 billion images, annotated with a class-hierarchy of around 30k labels via a semi-automatic pipeline" Epochs: 5M steps (Table 11) * 32768 (batch size) / 3B = 54.6 epochs

Epochs
54.6
Batch size
32,768

Training compute

The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.

Training compute
5.9 × 10²² FLOP

Digitizing Figure 9 indicates training used 27,200 TPUv3 core-days. TPUv3 is 123 teraflop/s per chip, 2 cores per chip. 1.23e14 * (1/2) * 27,200 * 24 * 3600 * 0.4 = 5.78e22 Alternatively, Table 2 indicates 965.3e9 FLOPs per forward pass on a 224^2 image. Table 4 indicates 5 million steps at a (normalized) batch size of 4096, and total flops including backward pass would be 3x the FLOPs from forward passes alone, so we get: 4096 * 5e6 * 965.3e9 * 3 = 5.93e22 (Note that actual batch size appear…

How it was established
Hardware,Operation counting

The training run

What it physically took to train: which chips, how many, for how long, and what that drew from the wall.

Training hardware
Google TPU v3
Chips used
2,048
Power draw
1.9 MW
Compute cost
$3,848

Availability

Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.

Weights
Closed — provider access only
Model access
Unreleased
Training code
Open source

About the weights: https://twitter.com/giffmana/status/1402507421029916672 About the code: Apache 2.0 https://github.com/google-research/big_vision/blob/main/big_vision/configs/proj/scaling_laws/train_vit_g.py

How it is classified

Labels the source dataset applies when tracking notable models, and how confident it is in the entry.

Why it is tracked
SOTA improvement

"we successfully train a ViT model with two billion parameters, which attains a new state-of-the-art on ImageNet of 90.45% top-1 accuracy"

Record confidence
Confident
Citations
1,393

Sources

Where this record came from and when it was last checked.

Reference
Scaling Vision Transformers
Last updated
25 May 2026

What the numbers mean

What this model is

ViT-G/14 was published by Google Brain,Google Research, in United States of America, in June 2021. The organisation is categorised as industry,Industry.

It works in Vision, and is recorded as doing image classification.

Because the weights are not available, none of the hardware figures elsewhere on this site apply to it.

What went into building it

Producing it required around 5.9 × 10²² FLOP of arithmetic, on Google TPU v3, which is a statement about the training budget rather than about inference.

It was trained on about 3,000,000,000 tokens of text.

The reason it appears in this catalogue at all is sOTA improvement.

Answers

ViT-G/14 — common questions

01

When was ViT-G/14 released?

ViT-G/14 was published in June 2021. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

02

What is ViT-G/14 used for?

ViT-G/14 works in Vision, and is recorded as handling image classification. A model can carry several of each, so these are the areas it was built for rather than a limit on what it will attempt.

03

How much compute was used to train ViT-G/14?

Around 5.9 × 10²² FLOP, on Google TPU v3. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.

04

What GPU do I need to run ViT-G/14?

None. ViT-G/14 is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.

05

Is ViT-G/14 open source?

No. ViT-G/14 has not had its weights published, so it exists only as a service controlled by its owner.

06

How many parameters does ViT-G/14 have?

ViT-G/14 has 1.8B parameters. Table 2 of paper. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.

07

Who created ViT-G/14?

ViT-G/14 was published by Google Brain,Google Research, based in United States of America, categorised as industry,Industry.

Source

Original publication

Record last updated 25 May 2026

The other direction

Looking at it from the other side?

This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.