What Is Apple Intelligence? Apple's AI Stack

Photo: Solen Feyissa / Pexels
Apple was late to generative AI. While Microsoft put Copilot into Windows and Google rebuilt products around Gemini, Apple said almost nothing in public until June 10, 2024, when it introduced Apple Intelligence at its developer conference. The name covers a layered system: a small model on your device, larger models on Apple-run servers, and an assistant (Siri) that sits on top. This article explains how the layers fit together, which devices run them, what shipped and when, and where the system still falls short.
The product in one paragraph
Apple Intelligence is a hybrid on-device and cloud AI system built into Apple's operating systems. Lightweight tasks run locally on the device's chip using a small Apple foundation model. Heavier tasks go to Private Cloud Compute (PCC), servers Apple designed so that personal data is used only to answer the request and is not stored or made accessible to Apple or anyone else, according to Apple's own description. For a time Apple also offered an optional hand-off to ChatGPT, introduced in December 2024, with a permission prompt.
Think of it as a hospital with three tiers. A nurse at the front desk (the on-device model) handles most routine questions instantly and never lets a file leave the building. A specialist in the next wing (PCC) takes the harder cases, but only inside a sealed room that is inspected from the outside. Anything beyond both would once have been referred to an outside consultant (ChatGPT), and only with your signature.
A short timeline, with dates
The product has changed enough since 2024 that a dated sequence is the clearest way to see it.
| Date | Event | Source |
|---|---|---|
| June 10, 2024 | Apple Intelligence announced for iOS 18, iPadOS 18, and macOS Sequoia | Apple Newsroom |
| October 28, 2024 | iOS 18.1 ships the first features (Writing Tools among them), in beta | MacRumors |
| December 2024 | iOS 18.2 adds Image Playground, Genmoji, and the optional ChatGPT hand-off | Apple release notes |
| March 2025 | Apple acknowledges the personalized Siri will take longer than planned | 9to5Mac |
| June 2025 | Foundation Models framework opens the on-device model to developers; Live Translation announced | Apple Newsroom |
| June 8, 2026 | Siri AI and the next generation of foundation models announced, built in collaboration with Google | Apple Newsroom |
| September 14, 2026 | iOS 27 and sibling releases ship; Siri AI begins rolling out in beta in English | Apple Newsroom |
The pattern is worth seeing in one glance: the privacy architecture and the writing and photo tools arrived on schedule, while the assistant that was meant to tie them together arrived about a year and a half after it was first demonstrated.
What the first wave actually did (2024 to 2025)
The first generation focused on features that make sense without an agent behind them.
- Writing Tools. System-wide proofreading, rewriting in a friendlier, more professional, or more concise tone, and summarizing. They work in Mail, Notes, Pages, and many third-party apps.
- Notification, Mail, and Messages summaries. Long threads collapsed into a few lines.
- Image Playground. Generates images from prompts in stylized looks. At the June 2024 announcement Apple described three styles (Animation, Illustration, and Sketch) and said images were created on device.
- Genmoji. Type a description and get a custom emoji-style image.
- Clean Up. Removes distracting objects from the background of a photo.
- Visual Intelligence. Point the camera at something and ask about it.
- Live Translation. Announced in June 2025, built into Messages, FaceTime, and Phone, and run on the device.
Notice what is missing from that list: anything that takes an action across apps on your behalf. That gap is the story of the next section.
For the underlying idea of small models doing useful work locally, see our explainer on LLMs versus SLMs and the older piece on small language models.
Why the smarter Siri took so long
At WWDC in 2024, Apple showed a Siri that would know your personal context, see what is on your screen, and act inside apps. In March 2025, Apple gave Daring Fireball a statement acknowledging that the more personalized Siri was delayed: "It's going to take us longer than we thought to deliver on these features," as reported by TechRadar and 9to5Mac.
Apple has not published a detailed post-mortem. The reliability problem is easy to grasp: an assistant that books, sends, and edits across apps has to be right nearly every time, because a wrong summary is an annoyance but a wrongly sent email is a mistake. We cover the general version of this problem in AI agents explained.
Apple's 2026 materials say the new generation of Apple Foundation Models was custom-built in collaboration with Google. The shifting dates suggest Apple was working to a quality bar rather than a calendar, though Apple has not said so.
Siri AI: what shipped on September 14, 2026
Apple released iOS 27, iPadOS 27, macOS 27, watchOS 27, and visionOS 27 on September 14, 2026, and said "Siri AI begins rolling out today in beta in English," with French, Japanese, Korean, Portuguese, and Spanish planned for the following month, per Apple's release post.
What Apple says it does:
- Personal context. It works across messages, emails, and photos, for example drafting an email or editing and sharing a photo.
- Onscreen awareness. It can act on what is displayed, such as adding a team's home games to a calendar.
- World knowledge. It draws on broad knowledge from across the web for up-to-date answers.
- A Siri app. Conversation history is kept in a dedicated app and synced through iCloud. Apple says the app "uses iCloud to privately sync conversational history across a user's products."
- Several ways to invoke it. "Hey Siri," the side button, or a swipe down from the Dynamic Island on iPhone; Spotlight on iPad and Mac; gaze on Vision Pro.
- Third-party apps. Existing app actions work, such as texting through WhatsApp or playing a book in Audible, with Microsoft Outlook, Notability, and Tripsy named as coming.
Apple says Siri AI orchestrates these actions through the Spotlight index and an "App Toolbox" that "work entirely on device." That detail matters for privacy: the system that decides which app to call runs locally, and only the model calls that need more power go to servers.
The limits Apple discloses
Three caveats appear in Apple's own release notes, and they are the parts most summaries skip.
- Daily usage limits. Features that rely on server-side models, including Siri AI, are subject to daily limits, and Apple says expanded access "will be available for a fee in the future."
- Regional gaps. Siri AI is not initially available in the EU on iOS, iPadOS, and watchOS, and not in China while Apple works through regulatory requirements. It is also unavailable to users under 13.
- Beta label. The English rollout is a beta, not a finished product.
Apple Intelligence itself has not carried a separate subscription price, but the fee language for expanded server access suggests that may change.
The models underneath
The company publishes more about its models than most people realize, and the numbers explain the hardware requirements.
Generation one and two: a roughly 3-billion-parameter model on the device
Apple's 2025 technical report describes "a roughly 3B-parameter on-device model" optimized for Apple silicon. Two techniques did the heavy lifting.
- KV-cache sharing. A transformer (the neural network design behind modern language models) stores "keys" and "values" for every word it has read so it need not recompute them. Per Apple's report, the on-device model shares this cache across parts of the network, which cuts memory use and time to first token. Think of two clerks sharing one set of notes instead of each keeping their own.
- 2-bit quantization-aware training. Quantization stores each model weight with fewer bits; training with that limit in mind keeps accuracy from collapsing. Our quantization explainer covers the general trade-off.
The 2025 server model used what Apple calls a Parallel-Track Mixture-of-Experts (PT-MoE) transformer, which combines track parallelism, sparse expert computation, and interleaved global and local attention. A mixture-of-experts design activates only some of its parameters for each token; see mixture of experts routing. Apple did not disclose the server model's parameter count.
Training data, per Apple, came from responsible web crawling, licensed corpora, and synthetic data. The Foundation Models framework, announced June 2025, gave developers Swift access to the on-device model with guided generation and tool calling, and inference is free of cost.
Generation three: five models
In June 2026 Apple described a third-generation family of five models in a research post.
| Model | Where it runs | What Apple says about it |
|---|---|---|
| AFM 3 Core | On device | Next generation of a 3-billion-parameter dense model |
| AFM 3 Core Advanced | On device (most capable chips only) | 20-billion-parameter sparse model activating 1 to 4 billion parameters per request; natively multimodal |
| AFM 3 Cloud | Private Cloud Compute | Optimized for speed and efficiency; size not disclosed |
| ADM 3 Cloud (Image) | Private Cloud Compute | Image generation and editing; size not disclosed |
| AFM 3 Cloud Pro | Private Cloud Compute | Most capable; agentic tool use and complex reasoning; size not disclosed |
The 20-billion-parameter on-device model is the surprising entry. A dense 20B model would not fit comfortably in a phone's memory, but a sparse design that activates only 1 to 4 billion parameters per request puts the working load in the range of a small model. It is a library with a small reading desk: the whole collection sits on the shelves, and only the books relevant to the question are carried over.
Apple's published evaluations are its own, using human side-by-side comparisons against its 2025 baselines, and should be read that way. For general text, AFM 3 Core was preferred on 45.6% of prompts against 23.3% for the 2025 on-device model, and AFM 3 Cloud was preferred on 64.7% against 8.7% for the 2025 server model. Apple also says it worked with Google and NVIDIA to extend Private Cloud Compute to NVIDIA GPUs in Google Cloud for AFM 3 Cloud Pro, keeping the same privacy guarantees. Apple says it does not use users' private personal data or interactions for training; a footnote updated September 9, 2026 adds that users can now opt in to help improve Siri and Apple Intelligence.
How Private Cloud Compute works
PCC is the part of the system Apple most deliberately differentiated. Apple's security research post lists five design requirements.
- Stateless computation. Personal data is used only to fulfill the request and is not retained after the response.
- Enforceable guarantees. The core protections do not depend on outside components such as load balancers.
- No privileged runtime access. There is no remote shell or debugging path, and privileged access cannot be expanded at runtime.
- Non-targetability. An attacker cannot single out one user's data without a broad, likely detectable compromise.
- Verifiable transparency. Researchers can check the guarantees and confirm production software matches what they inspected.
The hardware side uses custom Apple silicon with the Secure Enclave and Secure Boot, the same technologies found in iPhone. The Secure Enclave randomizes the data volume's encryption keys on every reboot and does not persist them, so data cannot survive a restart.
The transparency piece is the clever one. Measurements of all PCC code are published in an append-only, tamper-evident log, and Apple says production software images are published for inspection within 90 days of log inclusion. Your device sends data only to nodes whose attested measurements match a publicly logged release. This is like a restaurant that lets you read the health inspector's report for the exact kitchen your meal is prepared in, and your waiter refuses to carry the order if the report does not match. Apple also released a virtual research environment so security researchers can run a simulated PCC node on a Mac, and extended the Apple Security Bounty to the whole PCC stack.
The honest limit is that the guarantees rest on Apple's design and on researchers' ability to verify it, not on mathematical proof. For the general technique, see confidential computing for AI, and for the privacy and capability trade-off between local and cloud models, our piece on on-device versus cloud AI and the companion explainer on edge AI on your phone.
Which devices run it
Apple Intelligence has a hardware floor because the on-device model must live in memory beside the operating system. Apple's September 2026 requirements list:
| Category | Supported |
|---|---|
| iPhone | iPhone 16 or later, iPhone 15 Pro and Pro Max |
| iPad | iPad mini (A17 Pro); iPad with M1 or later |
| Mac | Mac with M1 or later; MacBook Neo (A18 Pro) |
| Other | Apple Vision Pro; Apple Watch Series 9 or later, Ultra 2 or later, or SE 3 when paired with a nearby Apple Intelligence iPhone |
The most capable on-device model, AFM Core Advanced, has stricter requirements. Apple lists iPhone Duo, iPhone 18 Pro models, iPhone Air, and iPhone 17 Pro models, iPad (M4) or later with 12 GB or more of memory, Mac (M3) or later with 12 GB or more of memory, or Vision Pro with M5. Those features include the expressive Siri voices and improved dictation.
Language support has widened since the initial launch, with some features limited to certain languages. Siri AI itself is English-first, with French, Japanese, Korean, Portuguese, and Spanish planned for the month after launch.
What developers get
The Foundation Models framework exposes the on-device model to third-party apps in Swift. Apple's June 2025 announcement said developers could use it in "as few as three lines of code," with guided generation (getting structured output) and tool calling built in, and with no per-request inference cost. That last point matters: a small app can add summarization or extraction without paying for API tokens.
The other developer surface is App Intents, the framework that lets apps expose actions to Siri and Shortcuts. In Apple's 2026 materials, outside apps are reached through the same App Toolbox that Siri AI uses to decide what to call. If you know the idea of function calling in language models, App Intents is the system-wide version; our explainer on function calling versus tool use covers the mechanics. The open-standard alternative, used by many non-Apple assistants, is described in What Is MCP?.
How it compares with other approaches
Specific competitor figures change quarterly, so the useful comparison is architectural.
| Question | Apple Intelligence | Typical cloud chatbot |
|---|---|---|
| Where does a simple request run? | On the device | On the provider's servers |
| Where does a hard request run? | Private Cloud Compute, with a public software log | Provider's data centers |
| Can the provider read the request? | Apple says no, by design | Depends on policy, not architecture |
| Who picks the model? | Apple | You, usually |
| Cost | No separate price so far; daily limits on server features | Free tiers plus paid plans |
Apple's differentiator is privacy architecture plus tight integration with the operating system, contacts, mail, and photos. The trade-off is that Apple controls the model choice and has historically trailed the frontier labs on raw capability. For a comparison of those frontier assistants, see ChatGPT vs Claude vs Gemini and What Is GPT-5?. Apple's own approach is now partly Google-derived, which blurs the older framing of Apple against Google.
What people get wrong about it
"It's just ChatGPT." The ChatGPT hand-off was optional and separate, with a permission prompt. Most Apple Intelligence features use Apple's own models. Apple's September 2026 release post does not mention ChatGPT, so check your device's settings for the current state.
"Everything runs on the phone." Only part of it does. Anything that escalates goes to PCC, which requires a connection.
"Private means perfect." Private means a smaller trust surface, not zero risk. Models can still produce wrong summaries and incorrect answers, a property shared by all language models; see why AI hallucinates facts.
"The 2024 demo shipped in 2024." It did not. The personal-context Siri was demonstrated in June 2024 and reached beta users on September 14, 2026.
Reading Apple's claims critically
Apple publishes more evidence than most consumer AI vendors, but the evidence has a shape worth understanding before you decide how much to trust it.
Benchmarks are Apple's own. The third-generation evaluations are human side-by-side preference tests run by Apple against Apple's own earlier models. A result like 64.7% preferred over the 2025 server model says the new model improved. It does not say how the model ranks against models from other labs, because those were not the comparison. For why leaderboard-style numbers deserve caution in general, see why AI benchmarks can be misleading.
Privacy claims split into two kinds. Some claims are architectural and checkable in principle: the transparency log, the attested software measurements, the absence of a remote shell. Others are statements of policy, such as the assertion that user data is not used for training. The first kind can be audited by outside researchers; the second depends on trusting the company. Both are useful, but they are not the same strength of guarantee.
"On device" and "private" are different words. A request handled locally never leaves the phone, which is the strongest privacy position. A request handled by PCC leaves the phone, encrypted to a specific verified server, and is protected by design rather than by distance. A request sent to any third-party service falls under that service's terms. Apple's earlier ChatGPT hand-off was the third case, which is why it carried its own permission prompt.
Usage limits are a quiet signal. Daily caps on server-based features tell you something about cost. Running large models on dedicated servers is expensive, and Apple's statement that expanded access will cost a fee in the future suggests the free tier is a launch position rather than a permanent one. That is an inference from Apple's wording, not an announced plan.
A practical way to check your own device
The facts above change with each software release, so the reliable check is on your own hardware rather than in any article. Open Settings, look for the Apple Intelligence and Siri section, and confirm three things: whether your device appears in the supported list, whether your Siri and device language is one of the supported languages, and which optional features, such as outside chatbot hand-offs, are switched on. If a feature you read about is missing, the usual causes are an unsupported device, an unsupported language, a region restriction, or a staged rollout that has not reached your account yet.
Frequently Asked Questions
Q: Which devices support Apple Intelligence? A: As of September 2026, iPhone 16 or later and iPhone 15 Pro and Pro Max, iPad mini (A17 Pro), iPads with M1 or later, Macs with M1 or later, Apple Vision Pro, and certain Apple Watch models paired with a supported iPhone. The most capable on-device model needs newer hardware with 12 GB or more of memory on iPad and Mac.
Q: Is Apple Intelligence free? A: There is no separate subscription price listed. Apple says server-based features, including Siri AI and Image Playground, are subject to daily usage limits and that expanded access will be available for a fee in the future.
Q: Does Apple Intelligence work offline? A: Features handled by the on-device model can run without a connection. Requests that escalate to Private Cloud Compute need internet access.
Q: How is Private Cloud Compute different from ordinary cloud AI? A: Apple designed it to be stateless, to lack remote administrative access, to publish its software measurements in a tamper-evident log, and to send data only to servers whose software matches that log. Ordinary cloud AI generally relies on the provider's policy rather than a verifiable architecture.
Q: What is Siri AI and when did it launch? A: Siri AI is Apple's rebuilt assistant with personal context, onscreen awareness, world knowledge, and a dedicated app. It began rolling out in English as a beta on September 14, 2026, with five more languages planned for October.
Q: Is Siri AI available in the EU or China? A: Not initially. Apple says it is unavailable on iOS, iPadOS, and watchOS in the EU at launch, and unavailable in China while it works through regulatory requirements.
Q: Does Apple use my data to train its models? A: Apple says it does not use private personal data or user interactions to train its foundation models, though users can now opt in to help improve Siri and Apple Intelligence. Training data comes from public, licensed, open-source, and synthetic sources.
The takeaway
Apple Intelligence is best understood as a privacy architecture with models attached, rather than the reverse. The on-device and Private Cloud Compute layers were in place by late 2024, and the part everyone was waiting for, an assistant that can use your context and apps, took until September 2026 and arrived as a limited English beta, with usage caps and gaps in the EU and China. Whether the privacy-first design holds up against faster-moving rivals will depend less on the keynote claims than on how Siri AI performs once daily use exposes its failures.
Sources
- Apple Newsroom, Introducing Apple Intelligence for iPhone, iPad, and Mac (June 10, 2024)
- Apple Newsroom, Apple Intelligence gets even more powerful (June 2025)
- Apple Newsroom, Siri AI, a profoundly more capable and personal assistant, is here (September 14, 2026)
- Apple Security Research, Private Cloud Compute
- Apple Machine Learning Research, Apple Intelligence Foundation Language Models Tech Report 2025
- Apple Machine Learning Research, Introducing the Third Generation of Apple's Foundation Models (2026)
- 9to5Mac, Apple Intelligence personal Siri delayed (March 2025)
Follow new articles
Add SciHub101 to your feed reader to see new science and technology stories.
About the Author
SciHub101 is an independent science and technology publication founded and edited by Buğra Sözeri, a software engineer and app developer based in Türkiye. AI tools assist with drafting and research. Source documentation in the existing archive is uneven, and articles are being reviewed and corrected. The Editorial Team byline identifies editorial responsibility; it does not certify specialist or independent human review. Where we compare products, anything built by our own team is labelled as such — see the editorial guidelines and About page for how corrections and disclosures work.
Related Articles

What Is Groq? Ultra-Fast LLM Inference
Groq's LPU runs open LLMs at 500-1000+ tokens per second, roughly 10x faster than typical GPU inference. Here's how and why it matters.

OpenAI o3 Explained — The Reasoning Model
OpenAI o3 explained: how the reasoning model works, what its ARC-AGI, GPQA and FrontierMath scores mean, what it costs, and where its limits are.
What Is MLX? Apple's ML Framework
MLX is Apple's open-source ML framework optimized for Apple Silicon. It exploits unified memory to run and fine-tune LLMs on Macs.


Chatbots Explained — From ELIZA to ChatGPT
60 years of chatbot history: from ELIZA's pattern matching to ChatGPT's 100M users. How modern conversational AI works, and what's still broken.

The Personal Computer: From Kit to Commodity — SciHub101
How the Altair 8800, Apple I, and IBM PC turned computing from an institution's tool into a personal one — and why IBM's open architecture was its biggest strategic blunder.