All-attention network + adaptive span

Closed weights Facebook AI Research 133M parameters July 2019

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
Facebook AI Research
Organisation type
Industry
Country
United States of America, France
Published
2 July 2019
Authors
Sainbayar Sukhbaatar, Edouard Grave, Guillaume Lample, Herve Jegou, Armand Joulin

What it does

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

Domain
Language
Task
Language modeling/generation

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
133M

133M (Table 3) For word level language modeling, we use a model with d = 512 and 36 layers, each with 8 heads and 2048 persistent vectors.

Training data
103,000,000 tokens

8k warmup steps A batch consists of 64 samples, each with 256 tokens Following Baevski and Auli [2] on WikiText-103, we use tied adaptive softmax and adaptive input with 3 clusters of size 20k, 40k and 200k assuming 200k steps: 200000*64*256/103000000 = 31.8 epochs

Epochs
31.8
Batch size
16,384

64*256

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
1.2 × 10¹⁹ FLOP

31330000000000 FLOP / sec / GPU [fp16] * 64 GPUs * 24 hours * 3600 sec/hour * 0.3 [assumed utilization] = 5.197271e+19 FLOP 6 FLOP / token / parameter * 200000 steps * 64 sequences per batch * 256 tokens per sequence * 133000000 parameters = 2.6148864e+18 FLOP sqrt(5.197271e+19*2.6148864e+18) = 1.1657733e+19 FLOP likely confidence: i am uncertain about number of steps and v100 precision _________ The Algorithmic Progress paper estimation was 4.6 × 10^19 FLOP

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
NVIDIA V100
Chips used
64
Wall-clock time
24 hours

"Training large models takes about a day on 64 V100 GPUs."

Power draw
39.5 kW

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
Unreleased

How it is classified

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

Record confidence
Likely
Citations
156
Benchmark data
All-attention network + adaptive span

Sources

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

Reference
Augmenting Self-attention with Persistent Memory
Last updated
25 May 2026

What the numbers mean

About this model

All-attention network + adaptive span was published by Facebook AI Research, in the country recorded as United States of America, during July 2019. It comes out of an organisation categorised as industry.

It works in the domain of Language, and is recorded as performing the task of language modeling/generation.

This is a closed model: the trained values stayed with whoever produced them, and there is no local version to run.

What went into building it

The training run consumed about 1.2 × 10¹⁹ FLOP, on hardware recorded as NVIDIA V100. That figure measures what producing the model cost, and has no bearing on how fast it answers.

It was trained on a corpus of about 103,000,000 tokens of text.

Answers

All-attention network + adaptive span — common questions

01

All-attention network + adaptive span— how many parameters does it have?

It has a parameter count of 133M. 133M (Table 3) For word level language modeling, we use a model with d = 512 and 36 layers, each with 8 heads and 2048 persistent vectors. 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.

02

All-attention network + adaptive span— who created it?

It was published by Facebook AI Research, based in United States of America, an organisation categorised as industry.

03

All-attention network + adaptive span— when was it released?

It was published in July 2019. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

04

All-attention network + adaptive span— what is it used for?

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

05

All-attention network + adaptive span— how much compute was used to train it?

Training consumed around 1.2 × 10¹⁹ FLOP, on hardware recorded as NVIDIA V100. 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.

06

All-attention network + adaptive span— what GPU do I need to run it?

None. This 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.

07

All-attention network + adaptive span— is it open source?

No. Its weights have not been published, so it exists only as a service controlled by its owner.

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.