FFN SwiGLU

Closed weights Google 220M 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
Google
Organisation type
Industry
Country
United States of America
Published
14 February 2020
Authors
Noam Shazeer

What it does

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

Domain
Language
Task
Language modeling, Question answering

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
220M

"We use the same code base, model architecture, and training task as the base model from [Raffel et al.,2019]. The encoder and decoder each consist of 12 layers, with dmodel = 768. For the attention layers, h = 12 and dk = dv = 64. The FFN layers have hidden size df f = 3072." [Raffel et al.,2019]: "Specifically, both the encoder and decoder consist of 12 blocks (each block comprising self-attention, optional encoder-decoder attention, and a feed-forward network). The feed-forward networks in e…

Training data
50,600,083,456 tokens

pre-training: "Identically to [Raffel et al., 2019], we pre-train for 524,288 steps on the span-filling objective on the C4 dataset. Each training batch consists of 128 examples, each of which has an input of 512 tokens and an output of 114 tokens" 524288*128*(512+114) = 42010148864 fine-tuning: "Fine-tuning consists of 131072 steps <..>, the input sequences for each step have a combined length of approximately 65,536 tokens." 131072*65536 = 8589934592 42010148864+8589934592 = 50600083456

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
3.4 × 10¹⁹ FLOP

6 FLOP/parameter/token * 220000000 parameters * 50600083456 tokens = 66792110161920000000 FLOP (10^19) 45000000000000 FLOP/GPU/sec * 21.85 hours * 3600 sec / hour * 16 GPUs * 0.3 [assumed utilization] = 16990560000000002000 FLOP (10^19) sqrt(66792110161920000000*16990560000000002000) = 3.3687317e+19

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
Google TPU v2
Chips used
16
Wall-clock time
22 hours

"Each training step took approximately 0.15 seconds on a 32-core TPUv2 cluster" -> 16 chips "we pre-train for 524,288 steps" 524288*0.15/3600 ~ 21.85 hours

Power draw
9.2 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.

Why it is tracked
Historical significance

the paper introduced SwiGLU

Record confidence
Confident

Sources

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

Reference
GLU Variants Improve Transformer
Last updated
28 November 2025

What the numbers mean

What this model is

FFN SwiGLU was published by Google, in the country recorded as United States of America, during February 2020. 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, Question answering.

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

What went into building it

The training run consumed about 3.4 × 10¹⁹ FLOP, on hardware recorded as Google TPU v2. That figure measures what producing the model cost, and has no bearing on how fast it answers.

Training consumed a corpus of around 50,600,083,456 tokens of text.

Its inclusion criterion: historical significance.

Answers

FFN SwiGLU — common questions

01

FFN SwiGLU— is it open source?

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

02

FFN SwiGLU— how many parameters does it have?

It has a parameter count of 220M. "We use the same code base, model architecture, and training task as the base model from [Raffel et al.,2019]. The encoder and decoder each consist of 12 layers, with dmodel = 768. For the attention layers, h = 12 and dk = dv = 64. The FFN layers have hidden size df f = 3072." [Raffel et al.,2019]: "Specifically, both the encoder and decoder consist of 12 blocks (each block comprising self-attention, optional encoder-decoder attention, and a feed-forward network). The feed-forward networks in each block consist of a dense layer with an output dimensionality of dff = 3072 followed by a ReLU nonlinearity and another dense layer. The “key” and “value” matrices of all attention mechanisms have an inner dimensionality of dkv = 64 and all attention mechanisms have 12 heads. All other sub-layers and embeddings have a dimensionality of dmodel = 768. In total, this results in a model with about 220 million 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.

03

FFN SwiGLU— who created it?

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

04

FFN SwiGLU— when was it released?

It 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.

05

FFN SwiGLU— what is it used for?

It works in the domain of Language, and is recorded as handling the task of language modeling, Question answering. 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.

06

FFN SwiGLU— how much compute was used to train it?

Training consumed around 3.4 × 10¹⁹ FLOP, on hardware recorded as Google TPU v2. 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.

07

FFN SwiGLU— 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.

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.