Apple Intelligence Explained: Siri AI, the AFM 3 Models, and the 2026 Next Generation in One Guide

Human Verified

Share to: , and

A calm, wide illustration of an iPhone, iPad and Mac on a desk with a soft glowing orb of light above them, one unified intelligence across devices. No text, no logos.

Apple Intelligence Explained: Siri AI, the AFM 3 Models, and the 2026 Next Generation in One Guide

Apple Intelligence is Apple’s answer to a simple question: where does AI belong on a device you already own? The answer is not a chatbot you open, and not a subscription you buy. It is a layer built into the operating system itself — writing help in Mail, summaries in Messages, photo editing in Photos, translation in a phone call, and a rebuilt Siri that can finally use what is on your screen and in your apps. It runs on your iPhone, iPad and Mac, it is free with a supported device, and since September 2026 it has a new generation behind it: Siri AI, a family of five new foundation models, and a much bigger cloud behind the ones that run on the device. Checked 20 November 2026 against Apple’s own pages, read in full.

Two things confuse almost everyone who meets Apple Intelligence for the first time, and this guide clears both up before anything else. The first is what it actually is. Apple Intelligence is not one product, and it is not Siri by another name. It is the name for the whole system: small models that run on your device, larger models that run on Apple’s servers, and features in a dozen apps that call on whichever model fits the job. The second confusion is the date. Most explainers written before September 2026 describe the old picture — the 2024 launch features and the old Siri — and that picture is now out of date. On 14 September 2026, iOS 27 shipped with the next generation: a rebuilt Siri called Siri AI, five new models, and new features in Photos, Safari, Messages and the Home app. Any guide that does not mention Siri AI is describing last year’s product.

There is also a correction that belongs at the top, because the loudest stories about Apple Intelligence in 2026 are wrong in an important way. Apple did announce a collaboration with Google in January 2026, and its new models were built with Google’s Gemini technology. But Apple is not running Gemini on your phone, and you cannot today swap Siri’s brain for Claude, ChatGPT or Grok. ChatGPT remains the only third-party model wired into the system, and the much-reported Extensions framework that would let you pick a different model exists in the code but has not been switched on. This guide says exactly what is live, what is beta, and what is still only a report — and it names the source for each.

A calm, wide illustration of an iPhone, iPad and Mac on a desk with a soft glowing orb of light above them, one unified intelligence across devices. No text, no logos.

What Apple Intelligence Actually Is — the Layer, Not an App

<img src="

An illustration of a thin glowing layer spread across phone, tablet and laptop screens, AI as a layer not an app. No text.

” alt=”An illustration of a thin glowing layer spread across phone, tablet and laptop screens, AI as a layer not an app. No text.”>

Start with the shape of the thing, because everything else follows from it. Apple Intelligence is not an app you download and not a chatbot you open when you need help. You will not find an Apple Intelligence icon on your home screen. Instead, the system lives inside the apps you already use, and it appears where the work is happening: a rewrite button in Mail, a summary at the top of a long Messages thread, a clean-up tool in Photos, a translation in the middle of a phone call. Apple’s own description, unchanged since the 2024 announcement, is that it is “a personal intelligence system” built into iPhone, iPad and Mac. The word “personal” does the heavy lifting: the system is supposed to know your context — your messages, your photos, your calendar — without that knowledge ever leaving your control.

That design choice explains why Apple Intelligence feels so different from ChatGPT or Gemini. A chatbot is a destination: you go to it, you type, you wait. Apple Intelligence is a layer: it comes to you, inside the app where the task already lives, and it tries to do a small job well rather than a large job impressively. Writing Tools does not write your novel; it tightens a paragraph you already wrote. Notification summaries do not read your mail for you; they compress a stack of alerts into one line so you can decide what matters. Clean Up does not restyle your photo; it removes the stranger who walked into the background. The ambition is deliberately narrow, and that narrowness is the point. Apple is betting that the AI people actually use every day is the one that saves thirty seconds inside a task they were already doing, not the one that asks them to learn a new habit.

Underneath the layer sit the models. There are two kinds, and the difference between them is the single most important technical fact about the system. The first kind runs on your device — small, fast models that live in the phone’s memory and work without any network connection at all. The second kind runs on Apple’s servers, for requests the on-device models cannot handle. Your device decides, request by request, which kind is needed, and the handoff is invisible. You do not choose; the system routes. The models themselves are called Apple Foundation Models, and the current generation — the third, known as AFM 3 — is the subject of its own section below, because it is where the 2026 story really lives.

There is one more structural fact worth fixing in your mind before the details. Apple Intelligence is free. There is no subscription for the core features, no per-request charge for the on-device models, and no Apple Intelligence tier in iCloud. What has changed in 2026 is that some server-side features now carry daily usage limits — Siri AI, advanced photo editing, Image Playground and the cloud models in Shortcuts among them — and Apple says expanded access will be available for a fee in the future. The WWDC26 announcement added that increased access comes with most iCloud+ plans. So the honest way to describe the price is: free to use, with a meter on the heaviest server features. A figure of $50 for a usage pack has been reported by this site’s own news desk, but Apple has not confirmed any price, and this guide does not quote it as an Apple price.

Finally, a word on what Apple Intelligence is not. It is not a general-purpose chatbot, and it will not pretend to be one. It has no long memory of your life beyond the personal context Siri AI can now search, it does not browse the open web for fun, and it does not generate long essays or write code for you. If you want a chatbot, Apple will happily hand you to ChatGPT — that handoff, and its limits, gets a full section later. What Apple Intelligence does is narrower and, for the people who own the devices, arguably more useful: it makes the apps you already live in slightly smarter, in the moment, without asking you to go anywhere else.

The 2026 Reset: Siri AI and the Next Generation Arrive

A calendar-like timeline illustration, a 2024 starting point and a brighter 2026 point, a reset and upgrade moment. No readable dates or words.

Apple Intelligence has had two lives, and the break between them is the most important date in this guide. The first life began on 10 June 2024, when Apple announced the system at WWDC24, and it reached the public on 28 October 2024 with iOS 18.1, iPadOS 18.1 and macOS Sequoia 15.1. That first release was deliberately small: system-wide Writing Tools, notification summaries, priority messages in Mail, and the Clean Up tool in Photos. Image Playground and Genmoji followed in December 2024 with iOS 18.2, along with the ChatGPT integration. Through 2025 the system grew steadily — more languages, Visual Intelligence on the iPhone 16’s Camera Control, Live Translation in iOS 26 — but the shape stayed the same, and Siri stayed the old Siri.

The second life began on 8 June 2026 at WWDC26, when Apple unveiled what it calls the next generation of Apple Intelligence and, with it, an entirely new version of Siri: Siri AI. The language in Apple’s announcement was unusually direct for a company that usually underplays. Siri AI, Apple said, is “profoundly more intelligent, knowledgeable, and capable” — a rebuilt assistant that understands personal context, sees what is on your screen, searches the web for current information, and takes actions across your apps. It comes with a dedicated Siri app, where you can revisit a past conversation or start a new one, and your conversational history syncs privately across your devices through iCloud. Developer testing began the same day, across iOS 27, iPadOS 27, macOS 27 and visionOS 27.

The public release came three months later, on Monday 14 September 2026, when iOS 27, iPadOS 27, macOS 27 Golden Gate, watchOS 27 and visionOS 27 shipped together. Siri AI arrived as a beta, in English only, and opt-in: after updating, you go to Settings, then Siri, then “Try Siri AI (Beta)”. At launch there was a waitlist, and waits reported by testers ranged from minutes to a couple of weeks; Apple later dropped the waitlist for most iPhone users. The beta label matters. Apple is explicit that Siri AI is still being finished, that some promised capabilities arrive later, and that the models will keep improving through the beta period. This guide treats every Siri AI claim as a beta claim, because that is what Apple says it is.

The same September release carried the rest of the next generation with it. Photos gained Spatial Reframing, an Extend tool and an improved Clean Up. Safari gained Notify Me, which watches a web page for changes — a restock, a price drop — and tells you when something moves. Image Playground was rebuilt around a new image model and can now produce photorealistic images. Messages gained one-tap suggestions based on the context of your conversation, and personalised Smart Reply. Mail gained suggestions and automatic proofreading. The Home app, the Journal app, the Finder and the accessibility features all gained their own intelligence, and the full list is in the sections that follow. None of this was a single feature launch. It was a new operating-system generation, and Apple Intelligence was the thread running through all of it.

The rollout has continued since September, and the dates matter because they are how you tell a current guide from a stale one. In October 2026, with iOS 27.2 on 26 October, Siri AI gained French, Japanese, Korean, Portuguese and Spanish. On 3 November 2026, the iOS 27.3 developer beta added four more Siri AI languages — Traditional Chinese, Norwegian, Turkish and Vietnamese — with a public release expected in December 2026, and opened Siri AI to every app developer through App Intents. In the same week, Apple Intelligence began beta testing in China, powered by Alibaba’s Qwen rather than Apple’s own models, and without Siri AI. All of that is less than a month old as this guide is checked, and it is the reason the guide is dated 20 November 2026 rather than pretending the story stands still.

The Five Models Behind It: Apple Foundation Models, Third Generation

Five glowing orbs of different sizes arranged in a row on a shelf, two small ones on a phone and three larger ones in a cloud, the five-model family. No text.

Every feature in this guide is powered by a model, and since June 2026 those models have names worth learning: Apple Foundation Models, third generation — AFM 3. Apple announced the family at WWDC26 and described it in detail in a research post on its Machine Learning Research site, updated on 9 September 2026. There are five models, split cleanly by where they run: two on your device, three on Apple’s servers. The split is the architecture, and once you see it, the whole system makes sense.

The two on-device models are AFM 3 Core and AFM 3 Core Advanced. AFM 3 Core is the next generation of the small dense model that has powered Apple Intelligence since 2024 — roughly 3 billion parameters, fast, and good at everyday language jobs like summarising, ranking and rewriting. AFM 3 Core Advanced is the headline: a 20-billion-parameter model that runs on a phone by never using all of itself at once. It uses a sparse architecture built on a technique Apple researchers call Instruction-Following Pruning. The full model lives in the phone’s flash storage, and for each request the system selects a small set of experts — activating only 1 to 4 billion parameters — and loads just those into memory. It is natively multimodal, which is what makes the expressive new voices and the much more accurate dictation possible, and Apple says it is “unlocked by and optimized for our most capable Apple silicon systems”, which is its way of saying the newest chips get the most from it.

The three server models run on Private Cloud Compute, Apple’s server system, and they are the ones that handle the jobs too big for a phone. AFM 3 Cloud is the workhorse: the standard server-side model, optimised for speed, efficiency and performance. ADM 3 Cloud is the image model — the one behind image generation and editing, the rebuilt Image Playground, Genmoji and the advanced photo tools like Spatial Reframing. AFM 3 Cloud Pro is the most capable model Apple has built: the one for agentic tool use and complex reasoning, the hardest requests in the system. Four of the five models are purpose-built for Apple silicon. Cloud Pro is the exception, and it is the most interesting fact in the whole family: for Cloud Pro, Apple worked with Google and NVIDIA to extend Private Cloud Compute onto NVIDIA GPUs running in Google Cloud, while keeping the same privacy guarantees. That sentence, from Apple’s own research post, is the clearest official statement of the Apple–Google relationship — the background on Google’s Gemini models, which Apple’s new models were built with, explains the other side of that partnership — and the next section of this guide comes back to it.

How good are these models? Apple publishes its own human evaluations, and they should be read as Apple’s own evaluations, not as independent benchmarks — Apple has not published a comparison against GPT, Claude or Gemini, and this guide does not invent one. On Apple’s side-by-side human tests for general text, the new AFM 3 Core was preferred on 45.6 percent of prompts against 23.3 percent for the 2025 model. The server model improved more dramatically still: AFM 3 Cloud was preferred on 64.7 percent of prompts against 8.7 percent for the 2025 server model, with roughly a 36 percent relative improvement in overall response satisfaction and a 21 percent relative improvement in instruction following. Cloud Pro adds roughly another 10 percent in overall satisfaction over Cloud, and 14 percent for image understanding and for maths. The voice numbers are the most concrete: on a five-point Mean Opinion Score, AFM 3 Core Advanced scored 4.15 for general voice and 4.24 for conversational voice, against 3.87 and 3.82 for Apple’s current production text-to-speech system — a gap Apple itself notes is highly noticeable, since a tenth of a point on this scale is already easy to hear. Dictation tells the same story: human graders preferred the new system on overall quality by 44.7 percent to 17.6 percent.

Two facts about how these models were trained deserve a paragraph, because they answer the question every privacy-minded reader asks next. Apple says it trains its foundation models on a mixture of publicly available information, data licensed or purchased from third parties, open-sourced data, data obtained through dedicated studies, and synthetic data. And it says, plainly, that it does not use users’ private personal data or user interactions to train its foundation models unless the user explicitly chooses to help improve them — an opt-in that Apple added in September 2026. Web publishers can also opt out of having their content used for training. None of that makes the models magically safe, and Apple runs multilingual post-training alignment, language-specific guardrail models and human red-teaming by native speakers across its supported locales under its Responsible AI principles. But the training-data line is a genuine difference from the way most large models are built, and it is worth stating exactly as Apple states it.

One more thing the models explain: why Apple Intelligence needs the hardware it needs. A 20-billion-parameter model that activates a few billion parameters per request, with weights shuttled between flash storage and memory, is not a job for any chip. It needs fast unified memory, a fast Neural Engine and a lot of storage for the models themselves — up to 14 GB on the newest devices. That hardware floor is the reason the device list in this guide is so short, and it is not a marketing decision. It is the models.

On-Device vs Private Cloud Compute: Where Your Request Actually Goes

A split illustration, a phone doing work locally on the left and a distant secure server hall on the right, a shield between them. No text.

The most important sentence in Apple’s description of its own system is a short one: “Apple Intelligence is aware of your personal information without collecting your personal information.” The mechanism behind that sentence is the split between on-device processing and Private Cloud Compute, and it is worth understanding properly, because it is the thing Apple is proudest of and the thing its competitors do not copy.

The rule is simple. When you ask for something, your device tries to do it itself first. Summarising a notification, rewriting a sentence, sorting your mail, cleaning up a photo — these are jobs for the on-device models, and they never leave the phone. Your data stays in your hand, the work happens without a network connection, and nothing is sent anywhere. Only when a request is too large or too complex for the on-device models does the system send anything to a server, and when it does, it sends it to Private Cloud Compute — Apple’s own cloud, built for this one purpose, and designed so that the privacy properties of on-device processing extend to the server.

Private Cloud Compute, announced with Apple Intelligence in June 2024, is a different kind of cloud from the ones that run other companies’ AI. It runs on custom Apple-silicon server hardware, with no persistent storage — the servers are, in Apple’s design, stateless. Your data is used only to answer your request and then erased; it is never stored, never logged in a way that could identify you, and never used to train models. The servers have no shell access and no way for an administrator to log in and look around. Before your device sends anything, it checks a cryptographic attestation proving that the server it is about to talk to is running a software image that appears in a public, append-only transparency log — software Apple publishes so that independent security researchers can inspect it. The request itself is encrypted end to end to that specific attested node, and it travels through a third-party relay using Oblivious HTTP, so Apple never sees your IP address and the relay never sees your content. Apple also provides a Virtual Research Environment that lets researchers run the Private Cloud Compute software on a Mac and test it, and it extended its security bounty programme to the system.

The 2026 change to this picture is the Cloud Pro tier. AFM 3 Cloud Pro, the most capable model, runs on NVIDIA GPUs hosted in Google Cloud — not on Apple’s own silicon, and not in Apple’s own data centres. Apple says it worked with Google and NVIDIA to extend Private Cloud Compute to that infrastructure “while maintaining the same guarantees to protect our users’ privacy”, and its security research site carries the details. This is the one place in the system where Apple’s privacy story depends on infrastructure it does not own, and it is the right place to be precise rather than reassuring. The guarantee is a cryptographic one — attestation, published software images, stateless computation — not a promise that no Google-owned hardware is involved. If you want the strongest version of Apple’s privacy, it is the on-device tier. If you want the most capable model, it runs on Google’s cloud under Apple’s attestation regime. Both facts are true at once, and a good guide says both.

None of this eliminates trust; it narrows it. That is Apple’s own framing, and it is the honest one. You still have to trust Apple’s hardware design, its silicon supply chain and the correctness of the attestation process. What Private Cloud Compute removes is the need to trust a policy document. The software the servers run is published, the measurements are in a public log, and your device refuses to talk to a server it cannot verify. That is a genuinely different design from a conventional cloud AI service, where your prompt is processed on infrastructure you cannot inspect under a privacy policy you cannot verify. Whether that difference is worth the hardware lock-in and the usage limits is a judgement for the reader — but the mechanism is real, and it is the reason Apple can make the sentence this section opened with.

Practically, there is one more thing to know about the split: it is invisible, and that is both a feature and a frustration. You cannot choose to keep a request on-device, and you cannot see which model answered you. Apple decides by routing each request to the smallest model that can handle it, and the only visible trace of the cloud is the daily usage limit. Server-side features — Siri AI, the advanced photo-editing tools, Image Playground and the cloud models in Shortcuts — are subject to daily limits that vary by feature, request complexity and system demand. Apple has not published the numbers, and this guide does not invent them. What Apple has said is that expanded access will be available for a fee in the future, and that increased access comes with most iCloud+ plans. The free tier is generous for everyday use; the meter is there for the heavy users.

Siri AI in Detail: Personal Context, Onscreen Awareness, App Actions

A friendly voice assistant orb above a desk scattered with messages, emails and photos, personal context. No readable words.

Siri AI is the reason most people will notice the 2026 generation, and it is also the part most likely to be misunderstood, because it is still a beta and because the Siri it replaces had fourteen years of low expectations to overcome. The old Siri was a command interpreter: you said a phrase it recognised, it did one thing, and anything it did not recognise became a web search. Siri AI is a different kind of system, and Apple’s announcement lists three capabilities that together define it: personal context understanding, onscreen awareness, and broad world knowledge with app actions. Each one deserves a plain explanation.

Personal context means Siri AI can search across your own stuff. Ask it when your mother’s flight lands, and it can look through your mail for the booking, your messages for the time she texted you, and your calendar for the pickup you scheduled — then combine them into one answer. Ask it for the recipe your friend sent you last month, and it can find the message rather than asking you which app it was in. The important detail is where this happens: Apple says the personal-context search runs on-device and on Private Cloud Compute, not by uploading your mail and messages to a server that keeps them. The context is yours, the search is local to your data, and the answer is generated from what it found. That is the capability Apple promised back in 2024 and then delayed; the 2026 version is the delivery, two years later, on new models that can actually do it.

Onscreen awareness means Siri AI can answer questions about what you are looking at. If a friend texts you an address, you can ask Siri to add it to your contacts without copying anything. If you are looking at a photo, you can ask who is in it, or when it was taken, or to send it to the person in it. The mechanism is that the system can read the content of the screen you are on — the text, the image, the app you are in — and use it as the context for your question. This is the feature that most clearly separates Siri AI from a voice search box: the question is short because the context is already there. “Send this to her” is a complete instruction when the system can see both the photo and the conversation.

App actions are the third leg, and the one that will matter most over time. Siri AI can take actions across apps — not just open them, but do multi-step work inside them, using the systemwide actions apps expose. Apple says the new Siri can “get things done across apps with even more systemwide app actions”, and the developer preview opened on 3 November 2026 is the mechanism by which every app can join in: developers describe their app’s actions through App Intents, and Siri AI can call on them. Apple is also introducing “skills”, which let developers teach Siri how to combine several actions into a single experience, and it is shipping three of its own skills in the first iOS 27.3 beta, including one that teaches Siri how to select and edit text in the Notes app. The customer release of the developer preview is, in Apple’s words, “in a few months”, with App Store submissions opening in spring 2027 — so the app-action story is still mostly a promise, and this guide marks it as one.

The dedicated Siri app is the visible sign of how different this is. For the first time, Siri has a home of its own: an app where you can revisit a past conversation or start a new one, in text as well as voice, with your conversational history synced privately across your devices through iCloud. The old Siri had no memory of a conversation once it ended; the new one keeps a thread you can return to. That sounds like a small change, and it is actually the largest change in how people will use it: a voice assistant you can type to, read back, and pick up where you left off is a different tool from a voice you summon and dismiss.

World knowledge is the last piece, and the one where Siri AI most resembles a modern chatbot. When a question needs current information, Siri AI can go out to the web, gather up-to-date information, and generate a helpful answer rather than handing you a list of links. Apple has published documentation for Applebot, the crawler that feeds Siri AI’s web answers, but it has not published rules about citations or source links, and neither of its newsroom announcements mentions them. So the honest statement is: Siri AI answers from the web, and Apple has not said how it shows its sources. If you need to check a Siri AI answer against its sources, that is a real limitation today.

The limits, stated plainly. Siri AI is a beta. It launched in English only, with French, Japanese, Korean, Portuguese and Spanish added in October 2026, and Traditional Chinese, Norwegian, Turkish and Vietnamese in the iOS 27.3 beta for a public release expected in December 2026. It is not available initially in the EU on iPhone, iPad or Apple Watch, though Mac and Apple Vision Pro users in the EU can access it when set to a supported language. It is not available in China at all. It is not available to users under 13. And the server-side features it depends on are subject to daily usage limits, with expanded access to cost a fee in the future. Apple also notes, in the fine print of its support page, that some promised capabilities — Apple names Live Rewind and Siri Recap in its reporting — arrive later. A beta that is honest about being a beta is better than a launch that pretends to be finished; Siri AI is the former, and this guide treats it as such.

What It Does in Your Apps: Writing, Mail, Messages and Notes

A desk with a letter, an inbox and a notebook, a pen rewriting text with a soft glow, writing help. No readable words.

If Siri AI is the headline, the everyday value of Apple Intelligence is in the apps — and the best way to understand it is to walk through a day with it switched on. The system meets you first in writing. Writing Tools works almost anywhere you can type: Mail, Messages, Notes, Pages, and third-party apps that use the standard text controls. Select a passage and you can rewrite it to sound more professional, more friendly or more concise; proofread it for grammar and spelling; or turn a long passage into a short summary, a list of key points, or a table. In iOS 27, Mail adds suggestions of its own, automatic proofreading catches errors as you type, and the Journal app offers customised writing prompts for entries. The point is not that Apple Intelligence writes for you. It is that it edits with you, in the place where the text already is, and the rewrite never leaves the app.

Mail is where the system earns its keep for most people, because mail is where volume hurts. Priority messages surface the emails that need you — the one from your child’s school, the boarding pass, the bill — above the newsletters and the notifications. Long threads can be summarised at the top, so you can see the decision without reading forty replies. Notification summaries do the same job for the lock screen: a stack of alerts from one app collapses into a single glanceable line. And the Reduce Interruptions Focus goes further, using intelligence to decide which notifications are worth breaking through and silencing the rest, with an Intelligent Breakthrough and Silencing system that learns what you treat as urgent. None of this reads your mail for a human. It reads it for you, on your device, and shows you less of it.

Messages gets the same treatment, with a few features that are specific to conversation. Missed a group chat with two hundred messages? A summary catches you up. Writing a reply? Personalised Smart Reply, new in iOS 27 and English-only at launch, suggests responses in your own style rather than generic ones. Planning something with friends? Intelligent poll suggestions can draft the poll from the context of the conversation — “Saturday or Sunday?” — in one tap. And Messages now offers one-tap suggestions based on the context of what you are talking about, so creating a reminder or a note from a conversation takes a single touch instead of a copy, a switch and a paste. You can also create a custom background for a conversation from a description, which is a small delight rather than a productivity feature, and it is fine for a system to have a few of those.

Notes is the quiet workhorse. Transcription summaries turn a recorded lecture or meeting into readable notes with the key points pulled out — the summary is available in English, French, German, Italian, Japanese, Korean, Portuguese, Spanish and Traditional Chinese, depending on your region. Image Wand, available on iPhone, iPad and Vision Pro, turns a rough sketch in a note into a finished illustration. And in the 2026 generation you can describe a reminder in natural language — “remind me to call the plumber when I get home” — and the system creates it, and describe a calendar event the same way. Shortcuts, too, now lets you describe a shortcut in natural language and use Apple Intelligence models inside your automations, which turns a feature most people never touched into one you can describe in a sentence.

The honest assessment of this layer is that it is genuinely useful and genuinely unglamorous, and that is a compliment. Nothing here will make a keynote audience gasp. What it will do is save you the five minutes of reading a thread you did not need to read, the embarrassment of a typo in an email to a client, and the friction of copying text between apps. The test of a system like this is not whether it can write a poem. It is whether, after a month, you miss it when it is off. For the writing and triage features, most people who have used them for a month would.

Photos and Creativity: Clean Up, Image Playground, Genmoji and Spatial Reframing

A photo being edited on a tablet, an unwanted object fading away, playful generated images floating beside it. No text.

Photos is where Apple Intelligence is most visible, because the results are pictures you can see. The tool most people meet first is Clean Up, which has been there since the 2024 launch and got meaningfully better in the 2026 generation. Tap an unwanted object — a stranger in the background, a bin at the edge of the frame — and the system removes it and fills in what should have been there. The improved version in iOS 27 is more accurate about edges and textures, and it works alongside two new tools: Spatial Reframing, which lets you improve the composition of a photo after it was taken by shifting the framing, and Extend, which expands the edges of a photo beyond what the camera captured. Together they amount to a quiet admission: the best photo is often the one you take a moment after the moment has passed, and the system can now fix that.

Search is the other half of the Photos story, and it is the half that saves the most time. Natural-language search lets you describe what you want — “photos of Maya at the beach last summer” — instead of scrolling through a library of fifty thousand images. In iOS 26 and later you can also create a memory movie from a description: type “our trip to the mountains, the golden light” and the system assembles a film from your library with music and pacing. It is a small feature that feels like a large one the first time it works, because it turns a library you never look at into something you actually watch.

Image Playground is the playful side, and the 2026 version is a genuine rebuild rather than a polish. The original, from December 2024, made cartoon-style images from a text description, in a handful of styles. The reimagined version, powered by the new ADM 3 Cloud image model, can create high-quality images in a photorealistic style as well, and it understands touch-based editing and personalisation — you can start from a photo of a person and generate images of them in different settings, and the results hold a likeness far better than the first version did. Image Playground is available on iPhone, iPad, Mac and Vision Pro, and it is one of the features subject to daily usage limits, because it depends on the server-side image model.

Genmoji is the smallest feature in this section and the one people use most often once they find it. Type a description — “a cat wearing a party hat” — and the system generates a custom emoji you can send in a message, styled to match the conversation. It is available across iPhone, iPad, Mac and Vision Pro, and it is the clearest example of Apple’s approach in miniature: a generative model doing one small, delightful job, inside the app where you were already typing, with no new habit to learn.

Two honest notes belong here. First, the image features are the most visibly limited by the daily usage caps, because image generation is the most expensive thing the server models do. If you generate dozens of images in an afternoon, you will hit the limit, and Apple has not published the number. Second, the photorealistic Image Playground raises the question every image generator now raises: what is real in your photo library when any image can be generated or extended? Apple does not answer that question in its product pages, and this guide does not pretend it does. The tools are labelled as generated in the apps that produce them, and the rest is a conversation the industry is still having.

Live Translation and Communication: Phone, FaceTime, Messages and AirPods

Two phones with speech bubbles in different colours flowing between them, live translation across languages, plus earbuds nearby. No readable words.

Live Translation is the feature that most clearly shows what on-device AI is for, because it is the one where latency matters most. A translation that arrives five seconds late is useless in a conversation; a translation that arrives as the other person speaks is a different world. Apple Intelligence translates live in three places: in Messages, where it translates the conversation as it happens; in Phone and FaceTime calls, where it translates spoken conversation in real time; and with AirPods, where it translates what the person in front of you is saying, in your ear, as they say it.

The language coverage is narrower than the full Apple Intelligence list, and it is worth stating exactly, because it is the detail most guides get wrong. Live Translation in Messages supports English, Danish, Dutch, French, German, Italian, Japanese, Korean, Norwegian, Portuguese, Spanish, Swedish, Chinese (simplified), Chinese (traditional), Turkish and Vietnamese. Live Translation in Phone and FaceTime is narrower still: English (US and UK), French (France), German, Italian, Japanese, Korean, Portuguese (Brazil), Spanish (Spain), and Chinese (Mandarin, simplified and traditional). Live Translation with AirPods works on AirPods 4 with Active Noise Cancellation and later, AirPods Pro 2 and later, and AirPods Max 2 with the latest firmware, paired with an Apple Intelligence-enabled iPhone, and it supports the same set as Phone and FaceTime. The pattern is consistent: text translation is broad, voice translation is narrower, and the languages arrive in waves rather than all at once.

Communication is also where the smaller features add up. Voicemail summaries condense a voicemail into a line you can read instead of a message you have to listen to — available in English, French, German, Japanese, Korean, Portuguese, Spanish and Traditional Chinese, with the exact countries listed on Apple’s support page. Call Context, new in iOS 27, brings intelligence to the phone call itself across iPhone, iPad, Mac and Apple Watch, though Apple says the least about this feature in its public pages and this guide does not overstate what it does. Enhanced order tracking in Wallet (English US and UK only) pulls delivery information out of your mail and shows it where you would look for it. And the Home app’s new intelligence — combining related activity notifications, describing what happened in a selection of HomeKit Secure Video clips before you watch them, and searching your clips by what happened in them — is communication of a different kind: your house telling you what it saw, in words.

The honest limit here is the same as everywhere else in this guide: availability varies by language and region, and a feature that works in English in the United States may not work in your language in your country. Apple’s support page is the authority for your combination, and this guide links it in the Sources rather than reproducing the whole matrix. Check before you buy a device for a single feature.

Which Devices Get It — and Which Ones Never Will

A row of devices on a table, some glowing and some dimmed, the supported versus unsupported divide. No text, no logos.

The device list is short, and it is the first thing most readers need. Apple Intelligence, in its current generation, runs on iPhone 15 Pro and iPhone 15 Pro Max, every iPhone 16 model and later — including the iPhone 17 Pro models, the iPhone Air and the foldable iPhone Duo announced on 9 September 2026 — on the iPad mini with the A17 Pro or A20 Pro chip, on iPad models with an M1 chip or later, on Macs with an M1 chip or later including the MacBook Neo with the A18 Pro, on Apple Vision Pro, and on Apple Watch Series 9 or later, Apple Watch Ultra 2 or later and Apple Watch SE 3 when paired with an Apple Intelligence-enabled iPhone. That is the whole list, from Apple’s support page, published 5 November 2026.

The exclusions are as important as the inclusions, because they are the source of most disappointment. The standard iPhone 15 and iPhone 15 Plus do not get Apple Intelligence, and no software update will change that. Every iPhone 14 and older is excluded. Every Intel Mac is excluded — Apple silicon is a hard requirement. Older iPads without an M-series chip are excluded, with the single exception of the iPad mini with the A17 Pro. If you own one of these devices, the features in this guide will not appear after any update, and any guide or video that tells you otherwise is wrong. The reason is not a marketing decision to make you upgrade, or at least not only that: the on-device models need 8 GB of memory and a Neural Engine fast enough to run them, and the newest models need up to 14 GB of storage for the model files alone. The hardware floor is a technical fact, and Apple has never suggested it will lower it.

Getting started is straightforward once the hardware is right. Update to the latest software — iOS 27, iPadOS 27, macOS 27 Golden Gate, watchOS 27 or visionOS 27 for the next generation, or at minimum iOS 18.1, iPadOS 18.1, macOS Sequoia 15.1, visionOS 2.4 or watchOS 11 for the first generation — and make sure Apple Intelligence is switched on in Settings under Apple Intelligence & Siri. Your device language and your Siri language must be set to the same supported language, or the features will not appear; if you change the Siri language, Apple Intelligence may stay unavailable until the new language finishes downloading. On an Apple Watch, the features come from the paired iPhone, and some need the iPhone nearby.

Storage deserves a paragraph of its own, because it surprises people. Apple Intelligence needs up to 8 GB of storage on most supported devices, and up to 14 GB on the newest ones — iPhone Duo, iPhone 17 Pro models and later, iPhone Air, iPads with an M4 chip or later and at least 12 GB of unified memory, Macs with an M3 chip or later and at least 12 GB, and the Apple Vision Pro with the M5 chip. Apple Watch needs up to 1.5 GB. Apple warns that storage used by the system may temporarily exceed those figures during and shortly after a software update, as new models download and old ones are removed. If your phone is full, that is the first thing to fix: the models cannot download, and the features will not appear, and the Settings screen will not always explain why. For a fuller walkthrough of the device list and the setup, the short explainer this site published in October 2026 covers the same ground in 829 words; this guide is the long version, and it links that piece for readers who want the quick answer.

The last point in this section is the one that decides purchases. Apple Intelligence is a reason to choose a supported device, but it is not a reason to replace a working one on its own. The features are genuinely useful, and Siri AI is a genuine step forward, but they are a layer on top of a phone you already use for calls, photos and messages — and a standard iPhone 15 still takes the same photos and runs the same apps. Buy the supported device when you were going to upgrade anyway. Do not upgrade for a summary feature.

Languages, Regions, and the EU and China Exceptions

A world map with some regions lit and others shaded, languages and borders. No text.

Apple Intelligence is available in most regions of the world, in sixteen languages, and that sentence is true and useless at the same time — because the exceptions are where the story is. The sixteen languages, as listed on Apple’s support page on 5 November 2026, are English, Danish, Dutch, French, German, Italian, Norwegian, Portuguese, Spanish, Swedish, Turkish, Vietnamese, Chinese (simplified), Chinese (traditional), Japanese and Korean. Vision Pro supports a smaller set: English, French, German, Italian, Simplified and Traditional Chinese, Japanese, Korean and Spanish. Some features work in fewer languages than the full list — personalised Smart Reply in Messages and Mail suggestions are English-only, for example, and Live Translation has its own narrower lists, given in the previous section. Your device language and Siri language must match, and they must both be one of the supported languages, or Apple Intelligence will not switch on.

Siri AI is narrower than Apple Intelligence, and the gap between the two lists is the single most confusing fact in this guide. Siri AI launched in English only. French, Japanese, Korean, Portuguese and Spanish arrived in October 2026 with iOS 27.2. Traditional Chinese, Norwegian, Turkish and Vietnamese are in the iOS 27.3 developer beta, with a public release expected in December 2026. German is an Apple Intelligence language, but it is not on the announced Siri AI expansion list — a detail worth stating, because German-speaking readers are the most likely to assume otherwise. Apple says it will “quickly expand support for more languages”, and the pattern so far is a new batch every few weeks rather than a single big release.

The European Union is the largest exception. The next generation of Apple Intelligence features is available in the EU on iPhone and iPad with iOS 27, and on Mac and Vision Pro — but Siri AI is not available initially in the EU on iPhone, iPad or Apple Watch. Mac and Vision Pro users in the EU can access Siri AI when set to a supported language. Apple attributes the delay to work around the EU’s Digital Markets Act, and the European Commission has publicly disputed that reading of the law, rejecting Apple’s proposed rollout timetable. This guide does not adjudicate that dispute; it reports the fact that matters to a reader in the EU, which is that the new Siri is not on their iPhone yet, and Apple has not given a date for when it will be. The earlier generation of Apple Intelligence features reached the EU in 2025 with iOS 18.4, so the pattern for the EU has been delay, not absence — but a pattern is not a promise.

China is a different story entirely, and it changed in November 2026. Until then, Apple Intelligence did not work on devices purchased in China mainland, and it did not work for devices bought elsewhere if you were in China mainland with a China Apple Account region. That changed with the iOS 27.3 developer beta on 3 November 2026, when Apple Intelligence began beta testing in China on iPhone, powered by Alibaba’s Qwen model rather than Apple’s own models, and extended to iPad and Mac with the second beta on 16 November. The China build offers a core set of the earlier features — Writing Tools, Genmoji, Live Translation, notification summaries and Visual Intelligence — and it explicitly excludes Siri AI and the other next-generation features from WWDC26. Apple announced on 31 October that iOS 27.3 would reach the public in December, and Shanghai’s cyberspace regulator confirmed on 9 November that Apple had completed the registration required for tablets and laptops. For the detail on the model behind the China launch, this site’s report on the Qwen-powered launch covers the announcement, and the background on Qwen itself, Alibaba’s model family, explains what Apple is borrowing.

The practical advice that follows from all of this is short. Before you count on any single feature, check Apple’s feature-availability page for your exact combination of device, language and region — it is the authority, and this guide links it in the Sources. Features arrive in waves, the waves are dated, and a guide that does not give you the dates is guessing.

What Still Needs ChatGPT — and the Extensions Question

A hand-off illustration, a phone passing a question to a larger chat bubble beyond it, the ChatGPT handoff. No readable words, no logos.

Here is the honest answer to the question in the original title of this guide: what still needs ChatGPT? Quite a lot — and less than the headlines suggest. The ChatGPT integration, called the ChatGPT extension, has been part of Apple Intelligence since iOS 18.2 in December 2024, and it is still there in iOS 27. When Siri AI meets a question it cannot answer from your context or the web, it can hand the question to ChatGPT. When Writing Tools needs to compose text from scratch rather than rewrite yours, it can use ChatGPT. Apple asks before sending anything, and you do not need an OpenAI account for the basic use — signing in is optional, and it is what unlocks more. The integration is available only in regions where the ChatGPT app and service is available, and on iOS 27 the switch moved to Settings, then Siri, then ChatGPT.

The reason ChatGPT is there at all is that Apple Intelligence is not a general-purpose chatbot and does not try to be one. It has no long general knowledge to draw on beyond what Siri AI can fetch from the web, and it is not built for open-ended conversation, brainstorming, or writing long text from a blank page. Those are exactly the jobs ChatGPT is built for, and the handoff is Apple’s way of admitting the gap without building a chatbot of its own. For the background on the model you are being handed to, the OpenAI explainer on this site covers the company and the model generations behind it.

That is the shipped reality. The reported future is more dramatic, and it needs careful handling, because it is the subject of the most misleading reporting about Apple in 2026. In May 2026, Bloomberg reported that the iOS 27 betas contain a framework called Extensions that would let users choose a third-party model — Claude, ChatGPT, Gemini or Grok — to power Siri, Writing Tools and Image Playground. The report was picked up widely, and some coverage described it as a shipped feature, a marketplace, or a done deal announced at WWDC. None of that is accurate. Apple did not announce Extensions at WWDC26 on 8 June, and it has not announced it since. As of 3 November 2026, careful re-checks found that no Claude, Gemini or Grok extension registers with Siri on any shipping iPhone, and ChatGPT remains the only third-party model a user can pick. The framework exists in the code; it is not switched on.

What would Extensions mean if it shipped? On the reporting, a user would install their preferred AI app from the App Store and choose it in Settings as the model behind Siri, Writing Tools or Image Playground — Claude for long writing, Gemini for image generation, ChatGPT for general questions, Grok for real-time context. Each provider’s privacy policy would govern the data routed to it, and Apple’s on-device tasks — summarisation, writing tools, image cleanup — would stay native. It would be the largest change to the iPhone’s AI since Apple Intelligence launched, and it would turn Apple from a model builder into a platform that hosts other companies’ models. That is why the reporting matters, and why the distinction between “in the code” and “switched on” matters more. For the model that would be the most likely alternative, the Claude models explainer on this site covers what Anthropic’s models do and where they are strong.

The Google question sits alongside this one, and it deserves the same care. In January 2026, Apple and Google announced a multi-year collaboration under which the next generation of Apple Foundation Models would be built on Google’s Gemini models and Google’s cloud technology, while Apple Intelligence kept running on Apple’s devices and Private Cloud Compute. Apple’s research post says the models are “custom-built in collaboration with Google” and that its security research describes using “the technology behind Google’s Gemini family”. Craig Federighi, in briefings after WWDC, is reported to have said that Apple is not running Gemini directly — that Gemini’s outputs were used to refine Apple’s own models in post-training, which is a distillation-style technique, not a handoff. Other descriptions say Siri is built with Gemini technology. The exact division is unsettled in public reporting, and the responsible statement is the one this guide makes: Apple built its models with Google’s help, the top server model runs on Google Cloud hardware, and Apple says the resulting models are its own. Anyone who tells you flatly that Siri “runs on Gemini” or that it “does not use Gemini at all” is overstating what is public.

So the answer to the title’s question, as of 20 November 2026: ChatGPT is still the only outside model Apple Intelligence can hand work to, and it handles the open-ended questions Apple’s own models are not built for. Extensions — the system that would let you pick Claude, Gemini or Grok instead — is built but not switched on, and no date has been announced. Watch the point releases: iOS 27.1 and 27.2 did not switch it on, and the next plausible window is a future update that Apple has not named.

Privacy: What Apple Promises, What It Can Prove, and the Limits

A shield and a locked vault with a magnifying glass, privacy that can be checked. No text.

Privacy is the first thing Apple talks about with Apple Intelligence, and the first thing a sceptical reader should test. Apple’s promises, in its own words, are threefold: powerful on-device processing, so that personal information is used without being collected; Private Cloud Compute, so that server requests are processed without your data being stored or accessible to Apple; and a training policy under which your private personal data is never used to train foundation models unless you explicitly opt in. The question is not whether Apple makes these promises — it does, prominently — but what backs them, and where they stop.

The strongest backing is architectural, and it is described in the Private Cloud Compute section of this guide: stateless servers, no persistent storage, cryptographic attestation against published software images, a public transparency log, encrypted requests through a third-party relay, and a Virtual Research Environment that lets independent researchers run the software and test it. This is the difference between a privacy policy and a privacy mechanism. A policy says “we will not look”. A mechanism makes looking hard, and publishes the code so that other people can check. Apple extended its security bounty programme to Private Cloud Compute, and the largest rewards are reserved for attacks that could expose request data. That is a meaningful commitment, because it invites the people most motivated to break the system to try.

The limits are real, and a good guide states them rather than letting the reader discover them. First, the attestation design means you are trusting Apple’s hardware, its silicon supply chain and the correctness of its attestation process. If any of those is compromised, the guarantees weaken — and the researchers who study confidential computing have shown that enclave technologies can be attacked with physical access to hardware. Private Cloud Compute protects you against a curious or compelled provider, not against someone inside the data centre. Second, the Cloud Pro tier runs on NVIDIA GPUs in Google Cloud, so the most capable model in the system depends on infrastructure Apple does not own; the guarantees there are cryptographic rather than physical. Third, the ChatGPT handoff sends data to OpenAI, under OpenAI’s policies, not Apple’s — Apple asks before sending, but once it is sent, Apple’s privacy regime no longer applies. Fourth, Siri AI’s web answers depend on Applebot’s crawl and on web sources Apple does not control, and Apple has not published how it cites them.

There are also practical controls worth knowing. Apple Intelligence is opt-in in the sense that you switch it on, and Siri AI is separately opt-in through its own beta toggle. You can turn off individual features — notification summaries, for instance — without turning off the whole system. You can opt in to help improve Siri and Apple Intelligence, which is what allows Apple to use some of your interactions for training; by default, you have not opted in, and the setting is yours to change. And the daily usage limits on server-side features are themselves a privacy-relevant fact: the meter exists because the server models are expensive to run, and the price of expanded access, when Apple announces it, will be the moment the free tier’s bargain becomes explicit. This site’s report on the Siri AI daily limits and the iCloud+ connection tracks that story as it develops.

The fair verdict on the privacy question is that Apple’s design is the most verifiable in consumer AI, and that verifiability is worth more than a promise — but it is not a proof that nothing can go wrong. The system is built to make misuse hard and detectable, not impossible. For most people, that is the right trade: a system that knows your context, keeps it on your device where it can, and sends the minimum necessary to a server it can prove things about, is a better deal than a chatbot that keeps everything. For people whose threat model includes Apple itself, or Google’s cloud, no design Apple ships will be enough, and this guide does not pretend otherwise.

For Developers: The Foundation Models Framework and Siri AI for Apps

A developer's workbench with building blocks and a small app window, apps tapping into on-device models. No readable words.

Apple Intelligence is not only a set of features Apple built. It is also a platform other developers can build on, and the developer story is where the system’s long-term importance lies. The mechanism is the Foundation Models framework, introduced in 2025 and expanded at WWDC26. It gives any app direct access to the on-device model — the same model that powers Apple’s own features — through a Swift API, with no API key, no network connection and no per-request cost. The app creates a language-model session, sends a prompt, and gets a response, all on the device. In the 2026 update the framework gained image input, so an app can pass a picture to the model and ask about it; Dynamic Profiles, which tune the model for a particular job; and a LanguageModel protocol that also accepts Private Cloud Compute, MLX and third-party providers, so the same code can reach a bigger model when the on-device one is not enough.

The practical consequence is that a small developer can add AI features to an app without running a server, paying per token, or sending user data anywhere. A journaling app can summarise entries on-device. A language app can generate practice sentences offline. A photo app can describe an image without uploading it. That is a genuinely different economics from the API model, where every request costs money and every request leaves the device. It is also, frankly, a strategic move: if the best AI features on iPhone are the ones that run on-device for free, developers have a reason to build for Apple’s platform rather than for a chatbot API.

The server side has its own developer path, and it is more gated. Since WWDC26, an app can swap the on-device model for a Private Cloud Compute model in the Foundation Models framework and send prompts and tool calls to Apple’s server model. That path is limited: it needs an entitlement Apple grants, it is available to developers in the App Store Small Business Program with fewer than two million first-time downloads, and every user gets a daily request limit. The gating is deliberate — server inference is expensive, and Apple is rationing it — but it means the most powerful models are not freely available to every developer, and the developers who want them have to ask.

The newest developer story is Siri AI itself. On 3 November 2026, with the iOS 27.3 developer beta, Apple opened Siri AI to every developer through App Intents. Any app can now tap into Siri AI’s personal-context understanding, onscreen awareness and natural language, and Apple introduced “skills” — a way for developers to teach Siri how to combine several actions into a single experience. Apple is shipping three of its own skills in the first beta, including one for selecting and editing text in Notes. The customer release is, in Apple’s words, “in a few months”, and App Store submissions open in spring 2027. This site’s report on Apple handing Siri’s keys to app makers covers the announcement and its limits, including that integrations built during the preview cannot be tested through external TestFlight or shipped in App Store releases yet, and that Siri AI is unavailable in the EU on iPhone, iPad and Apple Watch, so European developers cannot test it either.

For a reader who is not a developer, the takeaway is simple: the apps you use will get smarter in the next year, not because Apple adds features to them, but because their makers can now use Apple’s models. The first wave will be small and uneven — a summariser here, a smarter search there — and the quality will vary wildly. The second wave, when Siri AI opens to apps properly in 2027, is the one that could change how you use your phone: asking Siri to do something in an app you have never opened, and having it work, because the app taught Siri how. That is a promise, not a feature, and this guide marks it as one.

Verdict: Who Apple Intelligence Is For — and What to Watch Next

A crossroads with two signposts, one path leading to an Apple-style garden and one to a different road, the decision. No readable words.

So, after the models, the features, the device list and the exceptions — is Apple Intelligence any good? The honest answer is that it is very good at the job Apple designed it for, and that the job is narrower than the marketing suggests. If you own a supported iPhone, iPad or Mac, you already have it, it costs nothing, and the writing tools, the summaries, the photo cleanup and the live translation will save you time every week in ways you will stop noticing — which is the highest compliment a system layer can earn. Siri AI, in beta, is the first Siri that feels like it belongs in the same conversation as the modern assistants, and the personal-context search is the feature most likely to change how you use your phone once it is finished. The privacy design is the most verifiable in consumer AI, and that matters if you care where your data goes.

The case against is equally plain. Apple Intelligence is not a chatbot, and if you want one, you will be handed to ChatGPT — which is an admission, not a feature. The device list excludes the standard iPhone 15 and every Intel Mac, and no update will fix that. Siri AI is a beta with real gaps: no EU on iPhone, no China, limited languages, daily usage limits, and promised capabilities still to arrive. The Extensions framework that would let you choose your own model is built but not switched on, and Apple has said nothing about when it will be. And the most capable model in the system runs on Google Cloud hardware, which is a fact Apple states plainly and its critics will keep returning to. None of these are reasons to dismiss the system. They are reasons to describe it accurately, which is what this guide has tried to do.

Who should care most? If you write a lot of email, manage a busy inbox, take a lot of photos, or talk to people in other languages, Apple Intelligence is already worth switching on — the features are there, they work, and they are free. If you bought an iPhone 16 or later, or an M-series Mac, for the AI, Siri AI is the payoff you were waiting for, and the beta is worth joining. If you own a standard iPhone 15 or an Intel Mac, the honest advice is to wait for your next upgrade rather than to chase a workaround; there is none, and the ones sold online do not work. And if you are choosing between Apple Intelligence and a rival system — Microsoft’s Copilot, for example, which takes the opposite approach of putting the assistant front and centre rather than hiding it in the operating system — the difference is philosophical as much as technical. The Copilot explainer on this site lays out that rival approach, and the choice between an assistant you summon and intelligence that is simply there is the real decision.

What to watch next, in order. First, the iOS 27.3 public release expected in December 2026: four more Siri AI languages, and the first public Apple Intelligence in China. Second, the Siri AI developer preview maturing into a customer release “in a few months”, with App Store submissions in spring 2027 — the moment Siri AI stops being a beta and starts being a platform. Third, the Extensions question: whether Apple switches on the framework that lets you pick Claude, Gemini or Grok, and what it charges for it. Fourth, the price of expanded access: Apple has said it will cost a fee, and that most iCloud+ plans will include it, but it has not said how much. And fifth, the technical report Apple promised for summer 2026 on its foundation models, with updated evaluations and benchmarks — the document that will let independent readers judge AFM 3 rather than taking Apple’s word for it.

Apple Intelligence, checked 20 November 2026, is a system in the middle of its biggest change since it launched. The old Siri is gone, the new one is in beta, the models are new, the cloud is bigger, and the map of who gets what is still being drawn. This guide will be out of date by the spring — that is the nature of a system that ships with the operating system — but the structure it describes, the five models, the on-device and cloud split, and the privacy design underneath, are the parts that will last. If you remember one sentence, make it this one: Apple Intelligence is not a chatbot you open; it is a layer that comes to you, and the question is not whether it is clever, but whether it is there when you need it. On a supported device, in a supported language, in a supported region, it increasingly is.

Sources

  • Apple Support — “Get started with the next generation of Apple Intelligence” (support.apple.com/en-us/121115, published 5 November 2026): device list, storage requirements, 16 supported languages, full iOS 27 and iOS 26 feature lists, daily usage limits, EU and China rules, ChatGPT extension footnote. Read in full, 20 November 2026.
  • Apple Newsroom — “Apple unveils next generation of Apple Intelligence, Siri AI, and more” (8 June 2026, WWDC26): Siri AI announcement, AFM 3, feature list, availability, languages, iCloud+ note, performance claims from Apple’s April–May 2026 testing. Read in full, 20 November 2026.
  • Apple — “Apple Intelligence and Siri” (apple.com/apple-intelligence): overview page, privacy and Private Cloud Compute sections, Foundation Models framework for developers. Read 20 November 2026.
  • Apple Machine Learning Research — “Introducing the Third Generation of Apple’s Foundation Models” (machinelearning.apple.com, updated 9 September 2026): the five AFM 3 models, Instruction-Following Pruning, evaluation figures, Responsible AI principles, training-data policy, Google collaboration. Read in full, 20 November 2026.
  • Apple Newsroom — “Introducing Apple Intelligence, the personal intelligence system” (10 June 2024) and Apple Newsroom, 28 October 2024 (iOS 18.1 release): original announcement and first public release dates.
  • Apple and Google joint statement (12 January 2026), via secondary reporting: next-generation Apple Foundation Models built on Gemini models and Google cloud technology; Apple Intelligence to keep running on Apple devices and Private Cloud Compute.
  • Siri AI rollout reporting, 14 September – 3 November 2026 (secondary reads): 14 September 2026 beta launch in English; French, Japanese, Korean, Portuguese and Spanish in October 2026 (iOS 27.2, 26 October 2026); iOS 27.3 developer beta (3 November 2026) adding Traditional Chinese, Norwegian, Turkish and Vietnamese and opening the Siri AI developer preview via App Intents; waitlist later dropped for most users (site news post, 5018).
  • Apple Intelligence in China reporting, November 2026 (secondary reads): Qwen-powered beta from the iOS 27.3 developer beta (3 November 2026, iPhone; 16 November 2026, iPad and Mac); Siri AI excluded; Shanghai Cyberspace Administration registration confirmed 9 November 2026; public iOS 27.3 expected December 2026.
  • Extensions-framework reporting (Bloomberg, May 2026, via multiple secondary reads; re-checked 3 November 2026): framework present in iOS 27 betas; not announced at WWDC26; no Claude, Gemini or Grok extension registers on shipping devices as of 3 November 2026; ChatGPT remains the only third-party model.
  • Private Cloud Compute background: Apple’s 2024 announcement and Security Research materials, via secondary reads (2026): stateless computation, attestation, published software images, Virtual Research Environment, bug-bounty extension (reported up to $1,000,000).
  • Apple’s WWDC26 performance claims (apps up to 30% faster, photos up to 70% faster, AirDrop up to 80% faster): Apple’s own testing, April–May 2026, as footnoted in the 8 June 2026 newsroom release — not used in this guide.
  • Site sources linked in the article: the short Apple Intelligence explainer (post 313, 18 October 2026); the Qwen-powered China launch report (post 3934); the Siri AI daily-limits and iCloud+ report (post 6106); the Siri AI waitlist report (post 5018); the Siri developer-keys report (post 4380); the Gemini, Qwen, Claude, OpenAI and Microsoft Copilot explainers.

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.