MoE-Multi

Closed weights Jagiellonian University,Google Brain 8.7B parameters January 2017

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
Jagiellonian University,Google Brain
Organisation type
Academia,Industry
Country
Poland, United States of America
Published
23 January 2017
Authors
N Shazeer, A Mirhoseini, K Maziarz, A Davis

What it does

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

Domain
Language
Task
Language modeling, Translation
Numerical format
FP32

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
8.7B

Table 5 https://arxiv.org/abs/1701.06538

Training data
87,000,000,000 tokens

"We constructed a similar training set consisting of shuffled unique sentences from Google’s internal news corpus, totalling roughly 100 billion words" Assuming 100 words = 133 tokens

Epochs
10
Batch size
1,365,333

"Training was done synchronously on a cluster of up to 64 GPUs as described in section 3. Each training batch consisted of a set of sentence pairs containing roughly 16000 words per GPU." Although they appear to use word-level tokenization in other experiments, here they use subword tokens: "Similar to GNMT, to effectively deal with rare words, we used subword units (also known as “wordpieces") (Schuster & Nakajima, 2012) for inputs and outputs in our system." In total 64 GPUs * 16k words/GPU * …

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
9.4 × 10¹⁹ FLOP

12 days 64 NVIDIA K40 GPUs (see hardware data sheet for performance) 0.33 util rate

How it was established
Hardware

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
NVIDIA Tesla K40t
Chips used
64
Wall-clock time
288 hours (12 days)

12 days

Power draw
32.9 kW
Compute cost
$3,874

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

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
Highly cited,SOTA improvement,Historical significance

"On large language modeling and machine translation benchmarks, these models achieve significantly better results than state-of-the-art at lower computational cost"

Record confidence
Confident
Citations
4,587

Sources

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

Reference
Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer
Last updated
25 May 2026

What the numbers mean

About this model

MoE-Multi was published by Jagiellonian University,Google Brain, in Poland, in January 2017. The organisation is categorised as academia,Industry.

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

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

How it was trained

Training it took roughly 9.4 × 10¹⁹ FLOP of computation, on NVIDIA Tesla K40t — a measure of what producing the model cost, not of how fast it answers.

Around 87,000,000,000 tokens went into training it.

It is tracked in the underlying dataset for one reason in particular: highly cited,SOTA improvement,Historical significance.

Answers

MoE-Multi — common questions

01

What GPU do I need to run MoE-Multi?

None. MoE-Multi 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.

02

Is MoE-Multi open source?

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

03

How many parameters does MoE-Multi have?

MoE-Multi has 8.7B parameters. Table 5 https://arxiv.org/abs/1701.06538. 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.

04

Who created MoE-Multi?

MoE-Multi was published by Jagiellonian University,Google Brain, based in Poland, categorised as academia,Industry.

05

When was MoE-Multi released?

MoE-Multi was published in January 2017. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

06

What is MoE-Multi used for?

MoE-Multi works in Language, and is recorded as handling language modeling, Translation. These are the areas it was designed around; they describe intent rather than a hard boundary.

07

How much compute was used to train MoE-Multi?

Around 9.4 × 10¹⁹ FLOP, on NVIDIA Tesla K40t. 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.

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.