Feedback Transformer

Closed weights LORIA,University of Lorraine,Facebook AI Research 126M parameters February 2020

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
LORIA,University of Lorraine,Facebook AI Research
Organisation type
Academia,Academia,Industry
Country
France, United States of America
Published
21 February 2020
Authors
Angela Fan, Thibaut Lavril, Edouard Grave, Armand Joulin, Sainbayar Sukhbaatar

What it does

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

Domain
Language
Task
Language modeling

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
126M

Table 3 shows 126M. There is another instance of the Feedback Transformer mentioned in Table 9 with 139M parameters.

Training data
103,000,000 tokens

"The models are trained for 200k steps and the finetuned for additional 10k steps." from Table 9 (describes 139M model) batch size 512 sequence length 256 256*512*210000 / 103000000 = 267 epochs "speculative" confidence since there is no clear description of the 126M model and numbers are assumed based on 139M model description

Epochs
267.23
Batch size
131,072

256*512

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
7.7 × 10¹⁸ FLOP

6 FLOP / token / parameter * 126*10^6 parameters * 256 tokens per sequences * 512 sequences per batch * 210000 steps = 2.0808991e+19 FLOP assuming V100 GPU fp16: 31330000000000 FLOP/sec/GPU * 1 GPU * 84 hours * 3600 sec / hour * 0.3 [assumed utilization] = 2.8422576e+18 FLOP sqrt(2.0808991e+19*2.8422576e+18) = 7.690547e+18 FLOP ___________ in the Algorithmic progress paper they used estimation of 4.41e+19 FLOP also with low confidence

How it was established
Operation counting,Hardware

The training run

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

Wall-clock time
84 hours

" Our Feedback architecture takes 3.5 days to train" 3.5*24 = 84 hours

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.

Why it is tracked
SOTA improvement

"As shown in Table 4, the Feedback Transformer model achieves a new SOTA performance (on Enwiki8) of 0.96 bit-per-byte despite its small size."

Record confidence
Speculative
Citations
41
Benchmark data
Feedback Transformer

Sources

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

Reference
Addressing Some Limitations of Transformers with Feedback Memory
Last updated
28 November 2025

What the numbers mean

Background

Feedback Transformer was published by LORIA,University of Lorraine,Facebook AI Research, in France, in February 2020. academia,Academia,Industry is the category the publisher falls under.

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

Its weights were never published, so it can only be reached through its provider. No graphics card changes that.

How it was trained

The training run consumed about 7.7 × 10¹⁸ FLOP. That figure describes the cost of creating it and has no bearing on how quickly it generates text.

The training set ran to roughly 103,000,000 tokens.

Its inclusion criterion is sOTA improvement.

Answers

Feedback Transformer — common questions

01

What is Feedback Transformer used for?

Feedback Transformer works in Language, and is recorded as handling language modeling. 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.

02

How much compute was used to train Feedback Transformer?

Around 7.7 × 10¹⁸ 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.

03

What GPU do I need to run Feedback Transformer?

None. Feedback Transformer 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.

04

Is Feedback Transformer open source?

No. Feedback Transformer has not had its weights published, so it exists only as a service controlled by its owner.

05

How many parameters does Feedback Transformer have?

Feedback Transformer has 126M parameters. Table 3 shows 126M. There is another instance of the Feedback Transformer mentioned in Table 9 with 139M parameters. 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.

06

Who created Feedback Transformer?

Feedback Transformer was published by LORIA,University of Lorraine,Facebook AI Research, based in France, categorised as academia,Academia,Industry.

07

When was Feedback Transformer released?

Feedback Transformer was published in February 2020. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

Source

Original publication

Record last updated 28 November 2025

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.