AFM-on-device

Closed weights Apple 2.7B parameters July 2024

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
Apple
Organisation type
Industry
Country
United States of America
Published
29 July 2024
Authors
Andy Narayanan, Aonan Zhang, Bowen Zhang, Chen Chen, Chong Wang, Chung-Cheng Chiu, David Qiu, Deepak Gopinath, Dian Ang Yap, Dong Yin, Feng Nan, Floris Weers, Guoli Yin, Haoshuo Huang, Jianyu Wang, Jiarui Lu, John Peebles, Ke Ye, Mark Lee, Nan Du, Qibin Chen, Quentin Keunebroek, Ruoming Pang, Sam Wiseman, Syd Evans, Tao Lei, Tom Gunter, Vivek Rathod, Xiang Kong, Xianzhi Du, Yanghao Li, Yongqiang W…

What it does

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

Domain
Language
Task
Language modeling/generation
Approach
Self-supervised learning
Numerical format
BF16

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
2.7B

Table 1, sum of non-embedding and embedding parameters

Training data
7,588,000,000,000 tokens

Not explicitly mentioned, but I assume the 7.588T tokens do not involve multiple epochs.

Epochs
1
Batch size
18,949,753

Main pretraining uses sequence length of 4096 tokens; 4096 sequences per batch. During the "continued" pre-training stage, sequence length is upped to 8192 while batch size remains 4096. During context lengthening, sequence length is upped to 32768 while "the recipe is similar to that used for continued pre-training" implies same batch size of 4096. Weighting batch sizes by number of tokens seen in each stage: exp((6.3T * ln(4096 * 4096) + 1T * ln(8192 * 4096) + 100B * ln(32768 * 4096))/ 7.4T)…

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
4.5 × 10²³ FLOP

Model was initialized from a pruned version of a 6.4B parameter model trained using the same recipe as AFM-server. Assuming "same recipe" involves training for the full 6.3T tokens, this implies 6 * 6.3T * 6.4B = 2.42e23 FLOP. The pruning masks are learned by training over 188B tokens, which suggests 6 * 188B * 6.4B = 7.22e21 FLOPs. Pretraining is then run over 6.3T tokens; however, labels are a convex combination of true labels and the predicted labels from the unpruned 6.4B model. Since thi…

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
Google TPU v5p
Chips used
2,048
Hardware utilisation
MFU 52.0%

MFU is stated for AFM-server, assumed similar. If anything, this is likely a lower bound as training is done on a single slice of TPU chips compared to the 8 slices for AFM-server. "AFM-server was trained on 8192 TPUv4 chips provisioned as 8 × 1024 chip slices, where slices are connected together by the data-center network (DCN) [Chowdhery et al., 2022]. Only data-parallelism crosses the slice boundary, other types of state sharding are within-slice only as the within-slice interconnect bandwi…

Power draw
2.2 MW
Cloud vendor
Google Cloud
Data centre
Google us-east (either us-east1 or us-east5)

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
Hosted access (no API)
Training code
Unreleased

How it is classified

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

Likely above 10²³ FLOP
Yes
Why it is tracked
Significant use

Currently in beta access only, but will be integrated into millions or billions of iPhones.

Record confidence
Confident

Sources

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

Reference
Apple Intelligence Foundation Language Models
Last updated
28 November 2025

What the numbers mean

Background

AFM-on-device was published by Apple, in United States of America, in July 2024. The organisation is categorised as industry.

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

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

How it was trained

The training run consumed about 4.5 × 10²³ FLOP, on Google TPU v5p. That figure describes the cost of creating it and has no bearing on how quickly it generates text.

Around 7,588,000,000,000 tokens went into training it.

It is tracked in the underlying dataset for one reason in particular: significant use.

Answers

AFM-on-device — common questions

01

What GPU do I need to run AFM-on-device?

None. AFM-on-device 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 AFM-on-device open source?

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

03

How many parameters does AFM-on-device have?

AFM-on-device has 2.7B parameters. Table 1, sum of non-embedding and embedding 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.

04

Who created AFM-on-device?

AFM-on-device was published by Apple, based in United States of America, categorised as industry.

05

When was AFM-on-device released?

AFM-on-device was published in July 2024. 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 AFM-on-device used for?

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

07

How much compute was used to train AFM-on-device?

Around 4.5 × 10²³ FLOP, on Google TPU v5p. 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 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.