PermuteFormer

Closed weights Peking University 149.7M parameters September 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
Peking University
Organisation type
Academia
Country
China
Published
6 September 2021
Authors
Peng Chen

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
149.7M

Parameterization appears to be similar to a vanilla transformer. 6 layers, hidden dimension of 512, feed forward dimension of 1024, 8 attention heads. This would imply 20,447,232 parameters without embedding weights, and 512*vocab_size embedding weights (assuming tied embedding and unembedding projections) They appear to use word-level tokenization: "We evaluate unidirectional PermuteFormer on WikiText-103 (Merity et al., 2017). It is a language modeling dataset with about 103 million tokens," …

Training data
103,000,000 tokens

WikiText-103 is about 103M tokens

Epochs
30

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
2.8 × 10¹⁸ FLOP

6 * (30 * 103M) * 149,697,024 = 2.775e18 This seems a bit small relative to their statement: "It takes about 10 days on 8 V100 GPUs to get all the figures in this paper" which suggests about 2.7e20 FLOPs at 30% MFU. Table 2 indicates that Performer and PermuteFormer take 0.23x to 0.58x as long to train as a Transformer model. Figure 2 appears to be the most compute intensive figure, and would take about 4 * (2.775e18) + 1 * (2.775e18 / 0.365) = 1.9e19 FLOPs.

How it was established
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
8
Power draw
4.8 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
Open source

no specific license. training code: https://github.com/cpcp1998/PermuteFormer/tree/master/language_model

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

"Results show that PermuteFormer uniformly improves the performance of Performer, accelerates convergence, and achieves state-of-the-art on some tasks." "We conduct extensive experiments to evaluate PermuteFormer. It achieves strong empirical performance and obtains state-of-the-art on Long-Range Arena, a benchmark for efficient Transformers. It also improves the performance of Performer on language modeling tasks like WikiText-103"

Record confidence
Speculative
Citations
23
Benchmark data
PermuteFormer

Sources

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

Reference
PermuteFormer: Efficient Relative Position Encoding for Long Sequences
Last updated
25 May 2026

What the numbers mean

Background

PermuteFormer was published by Peking University, in China, in September 2021. It comes out of academia.

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

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 2.8 × 10¹⁸ FLOP, on NVIDIA V100. That figure describes the cost of creating it and has no bearing on how quickly it generates text.

Around 103,000,000 tokens went into training it.

It is tracked in the underlying dataset for one reason in particular: sOTA improvement.

Answers

PermuteFormer — common questions

01

How much compute was used to train PermuteFormer?

Around 2.8 × 10¹⁸ FLOP, on 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.

02

What GPU do I need to run PermuteFormer?

None. PermuteFormer 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.

03

Is PermuteFormer open source?

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

04

How many parameters does PermuteFormer have?

PermuteFormer has 149.7M parameters. Parameterization appears to be similar to a vanilla transformer. 6 layers, hidden dimension of 512, feed forward dimension of 1024, 8 attention heads. This would imply 20,447,232 parameters without embedding weights, and 512*vocab_size embedding weights (assuming tied embedding and unembedding projections) They appear to use word-level tokenization: "We evaluate unidirectional PermuteFormer on WikiText-103 (Merity et al., 2017). It is a language modeling dataset with about 103 million tokens," and I confirmed that word-level tokenization results in about 102M words across train-test-validation. If this is the case, there are 267,735 unique words, so the embedding layer alone would be 137,080,320 parameters, for a total of 149,697,024. 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.

05

Who created PermuteFormer?

PermuteFormer was published by Peking University, based in China, categorised as academia.

06

When was PermuteFormer released?

PermuteFormer was published in September 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.

07

What is PermuteFormer used for?

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

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.