ST-MoE

Closed weights Google,Google Brain,Google Research 269B parameters February 2022

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,Google Brain,Google Research
Organisation type
Industry,Industry,Industry
Country
United States of America
Published
17 February 2022
Authors
Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, William Fedus

What it does

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

Domain
Language
Task
Language modeling/generation
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
269B

269B. it's called ST-MoE-32B because it's equivalent to a 32B dense model.

Training data
1,500,000,000,000 tokens

"We pre-train for 1.5T tokens on a mixture of English-only C4 dataset (Raffel et al., 2019) and the dataset from GLaM (Du et al., 2021) summarized in Appendix E"

Batch size
1,000,000

" We use 1M tokens per batch"

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
2.9 × 10²³ FLOP

The paper claims "scaling a sparse model to 269B parameters, with a computational cost comparable to a 32B dense encoder-decoder". If this is true for training cost, then 6*32e9*1.5e12 = 2.9e23

How it was established
Operation counting

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

Apache License 2.0 Code for our models is available at https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/moe.py

How it is classified

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

Likely above 10²³ FLOP
Yes
Why it is tracked
SOTA improvement

"ST-MoE-32B improves the current state-of-the-art on the test server submissions for both ARC Easy (92.7 → 94.8) and ARC Challenge (81.4 → 86.5)."

Record confidence
Likely
Citations
375

Sources

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

Reference
ST-MoE: Designing Stable and Transferable Sparse Expert Models
Last updated
25 May 2026

What the numbers mean

Background

ST-MoE was published by Google,Google Brain,Google Research, in United States of America, in February 2022. industry,Industry,Industry is the category the publisher falls under.

It works in Language, and is recorded as doing language modeling/generation.

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

How it was trained

Producing it required around 2.9 × 10²³ FLOP of arithmetic, which is a statement about the training budget rather than about inference.

The training set ran to roughly 1,500,000,000,000 tokens.

Its inclusion criterion is sOTA improvement.

Answers

ST-MoE — common questions

01

When was ST-MoE released?

ST-MoE was published in February 2022. 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 ST-MoE used for?

ST-MoE works in Language, and is recorded as handling language modeling/generation. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.

03

How much compute was used to train ST-MoE?

Around 2.9 × 10²³ FLOP. 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 ST-MoE?

None. ST-MoE 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 ST-MoE open source?

No. ST-MoE has not had its weights published, so it exists only as a service controlled by its owner.

06

How many parameters does ST-MoE have?

ST-MoE has 269B parameters. 269B. it's called ST-MoE-32B because it's equivalent to a 32B dense model. 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 ST-MoE?

ST-MoE was published by Google,Google Brain,Google Research, based in United States of America, categorised as industry,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.