Apple has worked on machine learning for far longer than the phrase Apple Intelligence has existed. This timeline follows that story from Siri on the iPhone 4S in 2011 to the third generation of Apple Foundation Models and Siri AI in 2026.
Every entry has a date and links to an Apple source: Newsroom, developer.apple.com, Apple Support or Apple’s Machine Learning Research site. Milestones are product and platform moments. Research entries are Apple Machine Learning Research highlights, each with a short summary written by us.
You can filter by year and type, search the entries, and expand or collapse them. Every entry has its own link, so you can share a single moment.
2011
Siri launches on iPhone 4S
Apple announces iPhone 4S with Siri, an intelligent assistant you control by asking. This is the starting point of Apple's consumer AI story.
2017
Core ML brings on-device machine learning to developers
With iOS 11, Apple introduces Core ML, a framework that lets developers add machine learning to apps with all processing happening locally on the device.
Improving the Realism of Synthetic Images
Apple researchers show how to make simulator-generated images look more realistic, so that models trained on cheap synthetic data generalise better to real photos. It is an early example of Apple publishing its own computer-vision work.
Improving Neural Network Acoustic Models by Cross-bandwidth and Cross-lingual Initialization
The Siri team explains how it reuses data from existing acoustic models when Siri launches in a new language or handles a different audio bandwidth, because labelled speech data is scarce for every new launch.
Inverse Text Normalization as a Labeling Problem
Explains how Siri turns spoken words into tidy written forms, so that 'October twenty third twenty sixteen' appears as 'October 23, 2016'. The team treats it as a labelling problem and uses a compact statistical model.
Deep Learning for Siri’s Voice: On-device Deep Mixture Density Networks for Hybrid Unit Selection Synthesis
Describes how Siri's voice moved to deep learning from iOS 10 onwards, using on-device deep mixture density networks to pick and join speech units so the voice sounds smoother and more natural.
A11 Bionic brings the Neural Engine to iPhone
iPhone X ships with the A11 Bionic and a dual-core neural engine that Apple says performs up to 600 billion operations per second. It is built for machine-learning tasks such as Face ID.
Real-Time Recognition of Handwritten Chinese Characters Spanning a Large Inventory of 30,000 Characters
How Apple built real-time handwriting recognition for up to 30,000 Chinese characters that runs on iPhone, iPad and Apple Watch (Scribble mode), using deep learning on the device.
Hey Siri: An On-device DNN-powered Voice Trigger for Apple’s Personal Assistant
Apple explains how 'Hey Siri' works: a very small neural-network speech recogniser runs all the time, scores how likely it is that you said the phrase, and only then wakes the rest of Siri.
An On-device Deep Neural Network for Face Detection
Apple says it began using deep learning for face detection in iOS 10. The article covers the privacy and efficiency challenges of running that network on the device and the Vision framework that opened it to developers.
Learning with Privacy at Scale
Sets out the architecture Apple uses to learn from usage patterns, such as typing and websites visited, with local differential privacy so that individual data does not leave the device in readable form.
2018
Personalized Hey Siri
Recounts that 'Hey Siri' arrived with iPhone 6 (iOS 8) and describes how Apple adapts the trigger to the person speaking, so that it responds to the owner and is less likely to wake for others.
Finding Local Destinations with Siri’s Regionally Specific Language Models for Speech Recognition
Siri struggled with names of small local businesses. Apple describes adding regional knowledge to the speech-recognition language model so that local places are recognised more accurately.
A12 Bionic gets an eight-core Neural Engine
iPhone Xs and Xs Max launch with the first 7-nanometre smartphone chip, the A12 Bionic, with a next-generation eight-core Neural Engine.
Can Global Semantic Context Improve Neural Language Models?
Looks at word embeddings, the numerical maps of words behind typing, news and answers, and tests whether adding wider semantic context improves neural language models.
Optimizing Siri on HomePod in Far‑Field Settings
HomePod hears you from across a room, with echo, reverberation and noise. Apple describes the multichannel audio processing and speech work needed to make Siri reliable in that far-field setting.
2019
Bridging the Domain Gap for Neural Models
Neural networks often fail when real data differs from their training data. Apple presents a domain-adaptation method that uses task-specific decision boundaries to close that gap.
Language Identification from Very Short Strings
Explains how Apple's keyboard and language features decide which language a very short piece of text is in, which matters more as people type in several languages.
2020
M1 brings the Neural Engine to the Mac
M1 includes a 16-core Neural Engine that Apple says can perform 11 trillion operations per second, with up to 15x faster machine learning than the previous generation of Mac.
2021
Making Mobile Applications Accessible with Machine Learning
Apple describes using machine learning to make mobile apps usable for people with disabilities, for example by understanding what is on screen for people who are blind or have low vision.
Leveraging ML Compute for Accelerated Training on Mac
With M1 Macs and the ML Compute framework in macOS Big Sur, neural networks can be trained directly on the Mac. The article also points to the tensorflow-metal plug-in for GPU training.
Recognizing People in Photos Through Private On-Device Machine Learning
How the Photos app recognises people privately on the device: a pipeline of face and body models that powers curation, search and memories without sending the library to Apple.
Learning to Generate Radiance Fields of Indoor Scenes
A research paper on learning to generate 3D radiance fields of indoor scenes, building a model of what rooms tend to look like so new viewpoints can be predicted from few observations.
Improving Neural Network Subspaces
Explores why some minima that a neural network converges to generalise better than others, and how training subspaces of networks can improve accuracy.
On-device Panoptic Segmentation for Camera Using Transformers
Describes the transformer-based model that gives the Camera app pixel-level understanding of a scene, which supports features such as Portrait mode and semantic rendering.
Interpretable Adaptive Optimization
Statistical learning methods that help Apple teams measure which user experiences work best from noisy, privacy-protected feedback, including differential privacy and delayed data.
2022
Modeling the Impact of User Mobility on COVID-19 Infection Rates Over Time
A health-research paper that models how user mobility affected COVID-19 infection rates over time during the first phase of the pandemic.
Deploying Transformers on the Apple Neural Engine
Apple open-sources a reference PyTorch implementation of the Transformer architecture, tuned for the Apple Neural Engine, and explains how Transformers already power several Apple features.
A Multi-Task Neural Architecture for On-Device Scene Analysis
Instead of one model per feature, Apple describes a single multi-task network that produces shared scene-analysis signals used for photo search and Memories.
3D Parametric Room Representation with RoomPlan
How RoomPlan uses LiDAR and machine learning to turn a scan of a room into a parametric 3D model with walls, doors, windows and furniture.
2023
Improved Speech Recognition for People Who Stutter
Describes work on making speech recognition perform better for people who stutter, for both voice assistants and dictation.
Fast Class-Agnostic Salient Object Segmentation
Explains the on-device network behind the 2022 feature that lifts the subject out of a photo or removes the background in Photos, Preview, Safari, Keynote and other apps.
Learning Iconic Scenes with Differential Privacy
Shows how Apple uses differential privacy to learn which kinds of photos people take at popular places, used to choose key photos in Memories and Places in iOS 17, without personal data leaving the device.
Voice Trigger System for Siri
An overview of the voice-trigger system for Siri, designed to run entirely on the device with tight limits on power, memory and false triggers.
Advancing Speech Accessibility with Personal Voice
Describes Personal Voice, which lets people at risk of losing their speech, for example after an ALS diagnosis, create a synthesised voice of their own. Apple introduced it in May 2023 and shipped it in iOS 17.
2024
Deploying Attention-Based Vision Transformers to Apple Neural Engine
Explains how to deploy attention-based vision transformers on the Apple Neural Engine efficiently, building on the 2022 Transformers work.
Humanizing Word Error Rate for ASR Transcript Readability and Accessibility
Proposes a way to judge speech-to-text transcripts by how readable they are for people, not just by raw word error rate, motivated by Apple Podcasts transcripts and accessibility.
Enhancing Paragraph Generation with a Latent Language Diffusion Model
Presents a latent diffusion approach to generating paragraphs of text, as an alternative to autoregressive models that can become repetitive.
Personalizing Health and Fitness with Hybrid Modeling
A hybrid modelling approach that personalises heart-rate predictions from wearable data to a person's own context, such as activity.
Apple Intelligence announced at WWDC24
Apple introduces Apple Intelligence, built on a roughly 3-billion-parameter on-device model and a larger server model running on Private Cloud Compute.
Private Cloud Compute introduced
Apple describes PCC as a cloud intelligence system built with custom Apple silicon and a hardened operating system for private AI processing.
Improving How Machine Translations Handle Grammatical Gender Ambiguity
Explains how the Translate app lets users choose the grammatical gender of ambiguous words when translating from English into Spanish, French or Portuguese.
Combining Machine Learning and Homomorphic Encryption in the Apple Ecosystem
Describes how Apple combines machine learning with homomorphic encryption, for example private nearest-neighbour search for Enhanced Visual Search in Photos, to keep processing private.
First Apple Intelligence features ship
iOS 18.1, iPadOS 18.1 and macOS Sequoia 15.1 bring the first set of Apple Intelligence features to users.
On Device Llama 3.1 with Core ML
A technical walkthrough of optimising and running Llama 3.1 on Apple silicon with Core ML, for developers who want large language models on the device.
iOS 18.2 adds Image Playground, Genmoji, ChatGPT and visual intelligence
Apple Intelligence gains Image Playground, Genmoji, Writing Tools enhancements, ChatGPT support in Siri and visual intelligence. It also begins language expansion with localised English for Australia, Canada, Ireland, New Zealand, South Africa and the UK.
Accelerating LLM Inference on NVIDIA GPUs with ReDrafter
Apple brings its Recurrent Drafter (ReDrafter) speculative-decoding method to NVIDIA GPUs to speed up LLM inference for production use.
2025
Controlling Language and Diffusion Models by Transporting Activations
Introduces a technique to steer language and diffusion models by transporting their activations, giving finer control over outputs and helping reduce misuse.
Understanding Aggregate Trends for Apple Intelligence Using Differential Privacy
Explains how Apple uses differential privacy to understand aggregate usage trends in Apple Intelligence without seeing individual users' data.
An LLM-Based Approach to Review Summarization on the App Store
Describes the multi-step LLM system behind the review summaries on the App Store in iOS 18.4, designed to be inclusive, balanced and faithful to reviewers.
Foundation Models framework opens the on-device model to developers
At WWDC25 Apple introduces a new generation of foundation models designed to support 15 languages, plus a framework that gives developers direct access to the on-device model.
FastVLM: Efficient Vision Encoding for Vision Language Models
Presents FastVLM, a vision encoder designed to make vision-language models faster. Apple later released the models, including MLX checkpoints.
The “Super Weight:” How Even a Single Parameter can Determine a Large Language Model’s Behavior
Finds that a single parameter in a large language model, the 'super weight', can be essential: removing it makes the model produce gibberish.
Live Translation arrives with AirPods Pro 3
AirPods Pro 3 launch with Live Translation in beta for in-person conversations, powered by computational audio and Apple Intelligence.
Exploring LLMs with MLX and the Neural Accelerators in the M5 GPU
Shows how MLX takes advantage of the Neural Accelerators in the M5 chip on recent macOS beta, to run language models faster on a Mac.
2026
ParaRNN: Large-Scale Nonlinear RNNs, Trainable in Parallel
ParaRNN makes recurrent neural networks trainable in parallel, which for the first time makes it practical to scale them to billions of parameters.
WWDC26: next-generation Apple Intelligence and Siri AI
Apple previews iOS 27 and its siblings with the next generation of Apple Intelligence and Siri AI, described as an entirely new version of Siri.
Third-generation Apple Foundation Models
A family of five models built in collaboration with Google: AFM 3 Core (3B dense), AFM 3 Core Advanced (20B sparse, activating 1 to 4B parameters), AFM 3 Cloud, ADM 3 Cloud (Image) and AFM 3 Cloud Pro.
Private Cloud Compute expands to Google Cloud
Apple, Google and NVIDIA extend PCC to run the most demanding Apple Intelligence workloads on Google Cloud, with PCC privacy commitments kept.
Siri AI arrives in beta in English
Apple releases the next generation of Apple Intelligence with iOS 27 and its siblings. Siri AI rolls out in beta in English, with personal context, onscreen awareness and more app actions. It is not initially available in the EU on iOS, iPadOS and watchOS.
Siri AI adds French, Japanese, Korean, Portuguese and Spanish
With the 27.2 updates, Siri AI expands to five more languages. Apple says Chinese (Traditional), Norwegian, Turkish and Vietnamese will follow by the end of the year. Apple Intelligence itself supports 16 languages, including Danish, but Siri AI does not yet list Danish.
Faster Retrieval for Better, More Efficient RAG
Proposes faster maximum inner product search, the retrieval step at the core of recommendation and retrieval-augmented generation (RAG), so that large databases can be searched more cheaply.
No entries match.
How we built this
We use Apple’s own pages as the only sources. The list is not complete: Apple has published more than 1,100 research papers, and we have so far included the highlights Apple chose to feature. We add entries as Apple announces them. Siri AI language support, for example, is taken from Apple Support and changes over time, so check Apple’s page for the current list.