Citation and evidence

Apple Intelligence

36 min full readUpdated 38 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI Tools & ProductsArtificial Intelligence

Cite this article

Apple Intelligence is Apple's personal intelligence system, a suite of generative AI features built into iPhone, iPad, Mac, Apple Watch, and Apple Vision Pro that combines on-device machine learning models with a privacy-preserving cloud tier called Private Cloud Compute. Announced at Apple's Worldwide Developers Conference (WWDC) on June 10, 2024, and first shipped in October 2024, it powers writing assistance, image generation, an overhauled Siri, notification summaries, Live Translation, and a developer framework for running Apple's models locally. Apple markets it as "AI for the rest of us," with then-CEO Tim Cook saying it "will transform what users can do with our products."[1]

Apple Intelligence logo
Apple Intelligence logo. Source: Wikimedia Commons, Public domain.

As of June 2026, Apple Intelligence is available in more than a dozen languages across all supported iPhones, iPads, Macs, and the Apple Watch. The system launched with a limited feature set on October 28, 2024, and has expanded through a series of software updates. Its rollout drew sustained criticism: a SellCell survey conducted November 28 to December 6, 2024 found that 73% of Apple Intelligence users described the features as "not very valuable" or said they "add little to no value," and a class-action lawsuit over delayed Siri features produced a $250 million settlement in May 2026.[2][5]

What is Apple Intelligence?

Apple Intelligence is not a single app but a layer of intelligence woven into Apple's operating systems. It runs on a two-tier model architecture: a compact on-device model handles latency-sensitive and private tasks directly on the user's device, while heavier requests are routed to Apple's Private Cloud Compute servers, which run on custom Apple silicon and extend the privacy guarantees of on-device processing to the cloud. For the most demanding requests, the system can also call out to third-party models such as ChatGPT and Google Gemini, but only with explicit user permission. Apple says it does not use users' private personal data or interactions to train its foundation models unless they opt in; an opt-in to help improve Siri and Apple Intelligence was added in September 2026.[17][18][27]

History

WWDC 2024 announcement

Apple unveiled Apple Intelligence during its WWDC keynote on June 10, 2024, positioning it as "AI for the rest of us." The system was presented as a deeply integrated layer of intelligence woven into iOS 18, iPadOS 18, and macOS Sequoia rather than a standalone application. Craig Federighi, Apple's Senior Vice President of Software Engineering, demonstrated the system's capabilities across writing, image creation, and an overhauled Siri experience.[1]

The announcement came at a time when competitors like Google, Samsung, and Microsoft had already shipped consumer-facing AI features. Google's Gemini integration had been available on Pixel and Samsung devices for months, and Samsung's Galaxy AI suite had launched with the Galaxy S24 series in January 2024. Apple's approach differed by emphasizing privacy, on-device processing, and tight ecosystem integration over breadth of features.[2]

At the same keynote, Apple announced a partnership with OpenAI to integrate ChatGPT into Siri and writing tools, with the integration slated to arrive later in 2024. Apple CEO Tim Cook said the system "combines generative AI with a user's personal context to deliver truly helpful intelligence" while keeping that information private and secure.[1]

WWDC 2025

At WWDC 2025 on June 9, 2025, Apple expanded Apple Intelligence significantly with the iOS 26 / macOS Tahoe release. The company announced that Apple Intelligence would extend to Apple Watch for the first time via watchOS 26, enabling on-wrist AI features through features like Workout Buddy. The Foundation Models framework was announced as a new developer-facing API allowing third-party apps to run on-device language model inference directly. New end-user features announced included Live Translation for calls and messages, AI-powered call screening, Hold Assist, and voicemail summaries. Apple had said in March 2025 that the next-generation Siri capabilities, including personal context and in-app actions, would take longer than expected and would roll out "in the coming year."[13]

WWDC 2026

Apple's 37th annual WWDC ran June 8-12, 2026, at Apple Park in Cupertino, with the keynote on June 8. The headline announcement was Siri AI, a rebuilt assistant that can hold longer conversations, answer questions about what is on screen or in the camera view, generate text and images, and analyze files. A dedicated standalone Siri app debuted, and on iPhones with Dynamic Island the assistant's animations surface there during requests. In iOS 27, users swipe down from the middle of the screen to summon Siri AI. The redesigned voice supports customizable expressiveness and speech rate.[26]

Apple also introduced its third-generation foundation models, AFM 3, the engine behind these features. The new lineup comprises five models: two on-device (AFM 3 Core, a 3-billion-parameter dense model, and AFM 3 Core Advanced, a 20-billion-parameter sparse model that activates only 1 to 4 billion parameters per request) and three server-side (AFM 3 Cloud, AFM 3 Cloud (Image), and AFM 3 Cloud Pro). AFM 3 Cloud Pro is Apple's most capable server model, used for agentic tool use and complex reasoning.[27]

The most significant clarification concerned Google's role. Despite earlier reports describing a "custom Gemini" backbone, Apple said the whole AFM 3 family was "custom-built in collaboration with Google." At a post-keynote briefing, Apple AI vice president Amar Subramanya said the four Apple silicon models (AFM 3 Core, Core Advanced, Cloud and Cloud Image) were "refined using outputs from Gemini frontier models," a distillation approach rather than a wholesale adoption of Gemini, and described AFM 3 Cloud Pro's quality as "similar to Gemini frontier models." Craig Federighi said: "For these models, we use none of the models that Google deploys to their customers, nor do we use the infrastructure and means by which they deploy models to their customers."[32] AFM 3 Cloud Pro does run on Nvidia GPUs in Google Cloud: "For AFM 3 Cloud Pro, we worked with Google and NVIDIA to extend Private Cloud Compute to NVIDIA GPUs in Google Cloud, while maintaining the same guarantees to protect our users' privacy."[27][28]

Apple released iOS 27 and iPadOS 27 on September 14, 2026, with Siri AI and the dedicated Siri app included for Apple Intelligence-capable devices.[30] EU availability of Siri AI on iOS 27 and iPadOS 27 was again held back over the Digital Markets Act, with macOS 27 and visionOS 27 unaffected.[26]

Rollout timeline

Apple Intelligence did not ship as a single release. Apple rolled out features incrementally across multiple software updates over more than a year, a pace that drew criticism from users and press who felt the marketing of the iPhone 16 had overstated how complete the system was at launch.

ReleaseDateKey features added
iOS 18.1 / macOS Sequoia 15.1October 28, 2024Writing Tools, new Siri design, Type to Siri, notification summaries, Smart Reply, Clean Up in Photos, call recording and transcription, Priority Messages in Mail
iOS 18.2 / macOS Sequoia 15.2December 11, 2024Image Playground, Genmoji, Image Wand, ChatGPT integration with Siri, Visual Intelligence (iPhone 16 only), Writing Tools enhancements
iOS 18.3 / macOS Sequoia 15.3January 2025Notification summaries disabled for News and Entertainment apps, expanded Visual Intelligence capabilities (plant/animal identification, calendar event creation from posters)
iOS 18.4 / macOS Sequoia 15.4March 31, 2025Language expansion (Chinese (simplified), French, German, Italian, Japanese, Korean, Portuguese (Brazil), Spanish, and additional English locales), Priority Notifications improvements
iOS 26 / macOS TahoeSeptember 2025Live Translation, Workout Buddy (Apple Watch), call screening, Hold Assist, voicemail summaries, Foundation Models framework for developers, ChatGPT image generation styles
iOS 26.4Spring 2026Original target for next-generation Siri (personal context, on-screen awareness, in-app actions); Bloomberg reported in February 2026 that the features had slipped out of this release[34]
iOS 27September 14, 2026Siri AI, AFM 3 foundation models, redesigned Siri app, extended Visual Intelligence, broader third-party AI model choice (Claude, Gemini, ChatGPT)

Expanded article table

The $250 million class-action settlement reached in May 2026 (see Reception) reflects the legal exposure created by the gap between WWDC 2024 marketing and the phased reality.[5]

Features

Writing Tools

Writing Tools is a systemwide feature available wherever text can be edited, including Mail, Messages, Notes, Pages, and third-party applications. It provides three core functions:

  • Rewrite lets users generate alternative versions of their text with adjustable tone, such as making a message more professional, more concise, or more friendly.
  • Proofread checks grammar, word choice, and sentence structure, offering suggested edits with explanations the user can review or accept.
  • Summarize condenses selected text into a paragraph, bullet points, a table, or a list.

Writing Tools shipped with iOS 18.1 in October 2024 and received enhancements in iOS 18.2.[6]

Smart Reply

Smart Reply is available in Mail and Messages. It analyzes incoming messages and suggests contextually appropriate responses above the keyboard. In Mail, Smart Reply identifies questions within an email and ensures the suggested response addresses each one. The feature goes beyond generic quick replies by incorporating details about deadlines, availability, and other specifics referenced in the original message.[7]

Notification summaries

Notification summaries use on-device language models to condense long or stacked notifications into brief summaries on the Lock Screen. This is particularly useful for active group chats or message threads where many messages accumulate. The feature launched with iOS 18.1 but became controversial after generating inaccurate news headlines (discussed in the Reception section). Apple disabled summaries for News and Entertainment apps in iOS 18.3 (January 2025) while working on accuracy improvements, and brought them back in iOS 26 with a warning label.[8][38]

Priority Messages

Priority Messages appears as a dedicated section at the top of the Mail inbox, surfacing the most time-sensitive emails. The system identifies messages like same-day event invitations, boarding passes, or urgent requests and highlights them so users can act without scrolling through less important mail.[7]

Genmoji

Genmoji allows users to create custom emoji by typing a description into the emoji keyboard. The system generates multiple options, and users can further personalize results by referencing photos of friends or family members from their photo library. The generated Genmoji can be used inline in messages, shared as stickers, or added as Tapback reactions. Genmoji shipped with iOS 18.2 in December 2024.[9]

Image Playground

Image Playground is a dedicated image generation tool available both as a standalone app and integrated into apps like Messages. Users create images by providing text descriptions and selecting from styles including Animation and Illustration. Images are generated on-device, and with iOS 26, additional styles became available through ChatGPT integration, expanding the visual range considerably. Image Playground was released with iOS 18.2.[9]

Image Wand

Image Wand works with Apple Pencil on iPad within the Notes app. Users circle a rough sketch, and Image Wand transforms it into a polished image that matches the selected style. Circling an empty space prompts Image Wand to generate an image based on surrounding text and context in the note. It shipped alongside Image Playground in iOS 18.2.[9]

Clean Up in Photos

The Clean Up tool in the Photos app uses generative AI to remove unwanted objects and distractions from photographs. Users tap, brush over, or circle an element to remove, and the system fills the area with contextually appropriate content. The tool works best with smaller objects against uncluttered backgrounds. Clean Up was available from iOS 18.1.[10]

Visual Intelligence

Visual Intelligence is a camera-based AI feature initially exclusive to iPhone 16 models, activated by long-pressing the Camera Control button. Pointing the camera at an object, sign, plant, animal, or scene enables the system to summarize and copy visible text, translate text between languages, detect contact information, create calendar events from posters and flyers, and identify plants and animals. The interface includes an "Ask" button for ChatGPT queries and a "Search" button for Google image search.

Visual Intelligence debuted in iOS 18.2 and received expanded capabilities in iOS 18.3. By mid-2025 Apple had extended the feature to iPhone 15 Pro models, and iOS 26 extended it to on-screen content through the screenshot interface.[3] With iOS 27, screen and camera understanding became part of Siri AI itself, which can answer questions about what the camera is seeing.[11][26]

Call recording and transcription

Starting with iOS 18.1, users can record phone calls directly within the Phone app. When recording begins, all participants are automatically notified. After the call ends, Apple Intelligence generates a written transcription and summary.[6]

Live Translation

Introduced with iOS 26 in September 2025, Live Translation handles multilingual communication across Messages, FaceTime, and phone calls. In Messages, it automatically translates incoming texts into the user's preferred language. On FaceTime, it adds translated live captions to video calls. During regular phone calls, Live Translation speaks the translated version aloud so both parties hear the conversation in their native language. The feature works on-device for supported language pairs, maintaining privacy without routing audio to external servers.[3]

Call Screening

Also arriving with iOS 26, Call Screening uses Apple Intelligence to intercept calls from unknown numbers. When such a call comes in, the iPhone answers on the user's behalf, asks the caller to state their name and the reason for calling, and displays a live transcription. The user can then choose to pick up, send to voicemail, or block the number without ever speaking to the caller. This feature was noted by reviewers as a meaningful privacy and anti-spam tool, though some observed Apple was late to the capability relative to Google's Call Screen on Pixel phones, which has existed since 2018.[12]

Hold Assist

Hold Assist monitors hold music during phone calls and notifies the user when a human agent comes on the line. The iPhone detects when hold music has stopped and provides a notification, allowing users to set the phone down without the risk of missing when the call is answered. Hold Assist shipped with iOS 26.[3]

Voicemail summaries

Apple Intelligence generates brief, readable summaries of voicemail messages without requiring the user to listen to them in full. Summaries appear on the phone's notifications, letting users decide which voicemails require a callback. This feature also arrived with iOS 26.[3]

Workout Buddy

Workout Buddy is available on Apple Watch with watchOS 26, making Apple Intelligence available on the wrist for the first time. The feature draws on the user's workout history, current metrics, and fitness goals to deliver personalized motivational insights during exercise sessions. A text-to-speech model converts the insights into a generative voice modeled after Fitness+ trainers, calibrated to match the energy of the workout type. At launch, Workout Buddy supports Outdoor and Indoor Run, Outdoor and Indoor Walk, Outdoor Cycle, HIIT, and Functional and Traditional Strength Training, with English only at release.[3]

Siri rebuild

Apple Intelligence brought the most significant overhaul to Siri since its introduction in 2011, but the depth of the promised changes proved far harder to ship than Apple anticipated. The rebuilt assistant, branded Siri AI, was finally unveiled at WWDC 2026 and shipped with iOS 27 on September 14, 2026.[26][30]

Redesigned interface

Siri received a new visual design with a glowing light effect around the screen edges rather than the previous orb at the bottom. Users can interact with Siri via typing ("Type to Siri") in situations where speaking aloud is not practical. These changes shipped with iOS 18.1. With iOS 27, the interface evolved further: a standalone Siri app, card-based conversational responses, Dynamic Island animations, and a swipe-down gesture to summon the assistant.[6][26]

On-screen awareness

Apple demonstrated at WWDC 2024 that Siri would gain the ability to understand and act on content visible on screen. A user viewing a restaurant in Safari could ask Siri to add the address to a contact card; someone reading an article could ask Siri to summarize it. While this was among the most impressive demos from the keynote, the implementation was delayed. In March 2025, Apple formally confirmed that on-screen awareness would not ship in spring 2025 as originally targeted, and said the delayed Siri features would arrive "in the coming year." The capability ultimately surfaced as part of Siri AI in the iOS 27 cycle.[13][26]

Personal context

Apple described a version of Siri that could draw on personal data, including messages, emails, calendar events, contacts, and photos, to answer specific questions like "What time is Mom's flight landing?" or "Pull up the PDF that Jamie shared last week." This relies on on-device indexing and semantic search. Like on-screen awareness, personal context faced significant delays before arriving with the iOS 27 Siri overhaul.[13]

App actions and cross-app tasks

Apple promised Siri would perform actions within apps, not just launch them. Examples included sending a specific photo via Messages, editing a document in Pages, or rearranging items in a to-do list. Third-party developers can expose app actions through Apple's App Intents framework. The full scope of in-app actions was only partially delivered before 2026, with deeper cross-app handling tied to the broader Siri AI overhaul.[13]

Why was the Siri overhaul delayed?

The delay was not merely a matter of polish. Apple engineers hit structural problems. The previous version of Siri ran on two separate systems: an older platform handling basic tasks and a newer platform for advanced features. Merging them caused bugs. Apple's solution was to rebuild Siri from scratch using a "deeper end-to-end architecture," a fundamentally different approach that required substantially more time than a surface-level update.

Apple originally expected the next-generation Siri features to arrive in spring 2025. In March 2025, the company formally confirmed the delay to 2026. At WWDC 2025 in June 2025, Craig Federighi and Greg Joswiak acknowledged the delay, and the features were widely expected in spring 2026 with iOS 26.4; in February 2026 Bloomberg reported that testing problems had pushed them out of iOS 26.4. The rebuilt Siri AI was finally shown at WWDC 2026 on June 8, 2026, and shipped with iOS 27 on September 14, 2026.[4][13][26][30][34]

Lawsuit and $250 million settlement

The Siri delays triggered a class-action lawsuit against Apple, alleging false advertising and unfair competition. The complaint argued that Apple "promoted AI capabilities that did not exist at the time, do not exist now, and will not exist for two or more years," and that it had "saturated the internet, television, and other airwaves" with advertising for those features for months before pulling the ads when the delay was confirmed in March 2025. Apple agreed to settle in December 2025, and the $250 million settlement terms were made public on May 5, 2026. Eligible claimants are U.S. residents who purchased an Apple Intelligence-capable device (the full iPhone 16 range plus iPhone 15 Pro and 15 Pro Max) between June 10, 2024 and March 29, 2025, with payouts of $25 per device that could reach up to $95 per device if the volume of claims is low. Apple did not admit wrongdoing as part of the settlement.[5][29]

ChatGPT integration

Apple's partnership with OpenAI brought ChatGPT integration to Apple Intelligence starting with iOS 18.2 in December 2024. The integration operates in two primary ways: through Siri, which can forward complex queries to ChatGPT with user permission, and through Visual Intelligence on iPhone 16, where the "Ask" button sends captured images to ChatGPT for analysis. Users do not need a ChatGPT account for the basic integration, though a paid OpenAI account unlocks additional capabilities.

Apple's terms require that Siri explicitly ask the user before any information is sent to ChatGPT. The company also states that OpenAI does not store requests and that user IP addresses are obscured before transmission.[9]

With iOS 26 (September 2025), ChatGPT integration expanded to Image Playground, where it powers additional image generation styles beyond Apple's own Animation and Illustration styles.[3]

The ChatGPT integration continues alongside Apple's own AFM 3 models and the broader third-party model choice introduced in iOS 27, serving as one of several optional backends users can route specific requests to with explicit permission.[14][16]

Google partnership and the Gemini question

In fall 2025, Bloomberg reported that Apple was negotiating with Google to use a custom version of Gemini to power next-generation Siri, and Bloomberg later reported that Apple would pay about $1 billion a year for a custom Gemini model of roughly 1.2 trillion parameters. On January 12, 2026, CNBC reported that Apple had picked Google's Gemini to power the AI-upgraded Siri.[14][15][37]

When Apple detailed the arrangement at WWDC 2026, the picture proved more nuanced than the early reports suggested. Apple said all five AFM 3 models were "custom-built in collaboration with Google." The four models that run on Apple silicon were refined using outputs from Gemini frontier models, and Federighi said Apple uses "none of the models that Google deploys to their customers"; AppleInsider summarized the result as meaning users "never touch a drop of Google code, Gemini agents, or even Google Search."[32] The model runs on Nvidia GPUs hosted in Google Cloud, with Apple stating it "worked with Google and NVIDIA to extend Private Cloud Compute to NVIDIA GPUs in Google Cloud, while maintaining the same guarantees to protect our users' privacy."[27][28]

The Google relationship was widely interpreted as acknowledgment that Apple's own foundation models, while competitive for on-device tasks, needed reinforcement for the more ambitious Siri features Apple had demonstrated at WWDC 2024. Google gains a substantial new distribution channel and cloud-compute revenue; Apple gets access to leading frontier-model capabilities, via both distillation and infrastructure, without the multi-year timeline of training a comparable model entirely from scratch.[15]

iOS 27 third-party model selection

In May 2026, MacRumors reported that iOS 27 would allow users to select from multiple AI models for Apple Intelligence tasks, including Anthropic's Claude and Google Gemini, in addition to the existing ChatGPT option. This gives users more control over which AI backend handles their requests, potentially including different models for different use cases.[16]

Apple Private Cloud Compute

For tasks that exceed the on-device model's capacity, Apple Intelligence routes requests to Apple's Private Cloud Compute (PCC) infrastructure rather than a conventional cloud service. PCC runs on custom Apple-designed silicon servers and was built to extend the privacy properties of on-device processing to cloud workloads. At WWDC 2026, Apple extended PCC to Nvidia GPUs in Google Cloud for the AFM 3 Cloud Pro model while stating the same privacy guarantees continue to apply.[17][27]

The architecture uses end-to-end encryption from the user's device to validated PCC nodes. Supporting data center services such as load balancers and privacy gateways sit outside this trust boundary and cannot decrypt user requests. The Secure Enclave on PCC servers protects encryption keys, while Secure Boot and a Trusted Execution Monitor verify that only signed code runs. Attestation allows a user's device to cryptographically verify the configuration of a PCC cluster before sending any data.

Apple's stated guarantees for PCC:

  • User data is used only to fulfill the specific request and is not stored after the response is returned.
  • PCC nodes are not accessible to Apple employees, even those with administrative access.
  • The software running on PCC nodes can be audited by independent security researchers, who receive access to the code and the ability to verify what is running in production.

Apple published detailed technical documentation on PCC's architecture and invited external researchers to review it. Several security research firms have conducted audits.[17]

The 2025 server-side model running on PCC used a Parallel-Track Mixture-of-Experts (PT-MoE) architecture, combining track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention. Apple's 2024 technical report indicated the server model performed comparably to GPT-3.5-Turbo and Llama 3-70B on a range of tasks. AFM 3 Cloud, announced in June 2026, builds on that PT-MoE foundation.[18][27][33]

Foundation Models framework

At WWDC 2025, Apple announced the Foundation Models framework, a new developer API that allows third-party apps to access the on-device large language model at the core of Apple Intelligence.

The framework is available with iOS 26, iPadOS 26, and macOS 26 on any Apple Intelligence-compatible device where Apple Intelligence is enabled. It provides the approximately 3 billion parameter on-device model to developers through Swift APIs, enabling apps to run language model inference locally without any per-query cost, no internet connection required, and no user data leaving the device.

Key capabilities of the framework:

  • Guided generation allows developers to specify structured output schemas, constraining the model's responses to well-typed Swift data structures. This dramatically simplifies parsing model output in typed Swift applications.
  • Tool calling allows the model to invoke app-defined callbacks when it needs additional information to complete a request, enabling multi-step reasoning pipelines that integrate app data.
  • LoRA adapters let developers fine-tune the base model's behavior for domain-specific tasks by shipping small adapter weights alongside their apps rather than a full separate model.

Apple highlighted example use cases at WWDC 2025: generating personalized quiz questions from a student's study materials, producing summaries of workout performance data, and building context-aware writing suggestions within niche productivity apps. Because inference runs on-device, the framework is viable even in low-connectivity environments and does not require API keys or cloud accounts.[19]

At WWDC 2026 the framework was extended to the third-generation models, giving developers access to both AFM 3 Core (3B dense) and the larger AFM 3 Core Advanced (20B sparse) on capable devices.[27]

The Apple Foundation Models page covers the underlying model architecture, training methodology, and benchmark evaluations in detail.

Technical architecture

Apple Intelligence runs on a two-tier architecture: a compact on-device model for latency-sensitive tasks and a larger server-side model for more demanding workloads. The third generation announced in June 2026 (AFM 3) preserves this split while adding a larger sparse on-device model and a more capable cloud tier.

What models power Apple Intelligence?

Model (AFM 3, June 2026)TierParametersNotes
AFM 3 CoreOn-device~3 billion (dense)Next generation of the original on-device model
AFM 3 Core AdvancedOn-device20 billion (sparse)Activates 1-4 billion parameters per request; most powerful on-device model, natively multimodal
AFM 3 CloudServer (PCC)Not disclosedServer-side workhorse optimized for speed and efficiency
AFM 3 Cloud (Image)Server (PCC)Not disclosedImage generation and editing, powers the new Image Playground
AFM 3 Cloud ProServer (PCC)Not disclosedMost capable model; agentic tool use and complex reasoning; built with Google, runs on Nvidia GPUs in Google Cloud

Expanded article table

Apple described AFM 3 Core Advanced as storing the full model in flash memory (NAND) rather than DRAM, with a lightweight dense block selecting experts per prompt to enable inference-time elasticity.[27]

On-device model

The on-device component is an approximately 3 billion parameter language model optimized for Apple silicon. It uses a transformer architecture with grouped-query attention and shared input/output embedding tables to minimize memory footprint. Two key innovations enable it to run efficiently on mobile hardware:

TechniqueDescriptionImpact
KV-Cache SharingThe model is split into two blocks (62.5% and 37.5% of layers). Block 2 reuses key-value caches from Block 1.Reduces KV-cache memory usage by 37.5%
Low-bit weight compressionThe 2024 model used a mixed 2-bit and 4-bit configuration averaging 3.7 bits per weight; the 2025 update moved to 2-bit quantization-aware training.Shrinks the memory footprint for on-device inference

Expanded article table

In Apple's 2024 report the model reached about 0.6 milliseconds per prompt token time-to-first-token and about 30 tokens per second on iPhone 15 Pro; KV-cache sharing and 2-bit quantization-aware training were introduced in the 2025 update.[18][33]

Adapter system (LoRA)

Rather than maintaining separate fine-tuned models for each task, Apple uses LoRA (Low-Rank Adaptation) adapters. These are small neural network modules, typically tens of megabytes for the 3B model, that plug into the frozen base model's attention and feedforward layers. Different adapters handle different tasks (summarization, rewriting, code generation, etc.) and are swapped dynamically at inference time. Adapter weights use 16-bit precision with a rank of 16.[18]

Training methodology

Apple's foundation models are trained on the AXLearn framework (built on JAX and XLA) using a combination of data parallelism, tensor parallelism, sequence parallelism, and Fully Sharded Data Parallel (FSDP). Training data includes licensed content, web data crawled responsibly via AppleBot, and high-quality synthetic data. Post-training refinement uses supervised fine-tuning and reinforcement learning from human feedback (RLHF) with mirror descent policy optimization.

In human evaluations, Apple's 3B on-device model outperformed several larger open-source models on summarization and instruction-following tasks, including Phi-3-mini, Mistral-7B, Gemma-7B, and Llama 3-8B. The 2024 server model performed comparably to GPT-3.5-Turbo and Llama 3-70B. For the 2026 generation, Apple said its Apple silicon models were refined using outputs from Google's Gemini frontier models, while AFM 3 Cloud Pro was optimized for NVIDIA GPUs and runs in Google Cloud under Private Cloud Compute.[18][27][32]

Privacy approach

Apple structured Apple Intelligence around three tiers of processing, each with distinct privacy properties:

TierProcessing locationPrivacy guarantee
On-DeviceUser's iPhone, iPad, Mac, or Apple WatchData never leaves the device. All processing uses the local model (AFM 3 Core / Core Advanced) and LoRA adapters.
Private Cloud ComputeApple silicon servers, plus Nvidia GPUs in Google Cloud for AFM 3 Cloud ProData is encrypted in transit, used only for the specific request, deleted immediately after response, and not accessible to Apple staff. Code is auditable by independent researchers.
Third-Party ModelsExternal servers (OpenAI for ChatGPT, optional Gemini/Claude in iOS 27)Explicit user permission required before any data is sent. Apple states that third-party providers do not store requests. IP addresses are obscured.

Expanded article table

Apple says users' private personal data is not used to train its foundation models unless they opt in (an opt-in was added in September 2026).[27] The PCC architecture was designed to extend the security guarantees of on-device processing to cloud workloads, and Apple has invited independent security researchers to verify these claims by inspecting the code running on PCC servers.[17]

Supported devices

Apple Intelligence requires hardware with at least 8 GB of RAM and a sufficiently capable Neural Engine.

Device categoryCompatible modelsChip requirement
iPhoneiPhone 15 Pro, iPhone 15 Pro Max, all iPhone 16 models, all iPhone 17 models (including iPhone 17e), all iPhone 18 models[30]A17 Pro or newer
iPadiPad Air (M1 and later), iPad Pro (M1 and later), iPad mini (A17 Pro and later)M1 or A17 Pro or newer
MacAll Macs with Apple siliconM1 or newer
Apple WatchApple Watch Series 10 or later, Apple Watch Ultra 2 or later and Apple Watch SE 3 (watchOS 27); Workout Buddy on watchOS 26 also ran on Series 6 or later, SE (2nd generation) and Ultra models[3][26]Paired Apple Intelligence-capable iPhone nearby

Expanded article table

The iPhone 15 and iPhone 15 Plus do not support Apple Intelligence despite running iOS 18/iOS 26 because their A16 Bionic chip has only 6 GB of RAM. Apple executives pointed to the A17 Pro's 16-core Neural Engine, which performs nearly 35 trillion operations per second, and hinted that RAM also matters; every compatible device has at least 8 GB.[20]

Geographic rollout

United States

Apple Intelligence launched in U.S. English on October 28, 2024, with iOS 18.1. The initial release required the device and Siri language to be set to U.S. English but worked in most regions, with localized English for other countries following in December.[6]

European Union

Apple withheld Apple Intelligence from EU countries at the iOS 18.1 launch, citing concerns about compliance with the Digital Markets Act (DMA). Apple stated that the DMA's interoperability requirements could force changes that "compromise the integrity of our products in ways that risk user privacy and data security." The features delayed for EU users included Apple Intelligence, iPhone Mirroring, and SharePlay enhancements.

Apple Intelligence eventually launched in the EU in March 2025, alongside expanded language support. The episode prompted criticism from EU regulators and consumer groups who argued Apple was using the DMA as a pretext to delay competition-enabling features. Apple denied this framing. The DMA tension recurred at WWDC 2026, when Apple again held Siri AI on iOS 27 and iPadOS 27 back from the EU at launch.[21][26]

Additional English locales and languages

Localized support for additional English variants (UK, Australia, Canada, etc.) arrived in December 2024 with iOS 18.2. Broad language support, covering Chinese (simplified), French, German, Italian, Japanese, Korean, Portuguese (Brazil) and Spanish, arrived in March 2025 with iOS 18.4, roughly five months after the U.S. launch. Vietnamese was among eight further languages Apple announced at WWDC 2025 for later that year.[3]

China

China presented the most complex rollout challenge. Apple's AI model had to comply with Chinese content regulations managed by the Cyberspace Administration of China (CAC), which requires that generative AI systems pass a security evaluation and complete algorithm filing before reaching users. Foreign AI models are not permitted under this framework, so Apple needed domestic AI partners.

Apple struck a deal with Alibaba Group in early 2025 to use Alibaba's Qwen large language model as the AI backend for Apple Intelligence in China. A separate arrangement with Baidu for Visual Intelligence features was also reported. As of late 2025, the technology was reportedly ready but blocked pending CAC regulatory approval. CEO Tim Cook addressed the delay in October 2025 during a visit to Shanghai, confirming Apple was actively working toward a Chinese launch.

In a notable incident on March 30, 2026, Apple Intelligence briefly appeared on Chinese iPhones without regulatory clearance, after appearing with the first iOS 26.5 developer beta; it was gone again the same day. Full regulatory approval had not been granted as of mid-2026. On October 30, 2026, Apple said iOS 27.3, due in December 2026, would enable some earlier Apple Intelligence features on iPhone 15 Pro and newer models in China, with Alibaba's Qwen supported in Siri, Writing Tools and Visual Intelligence; Siri AI and the WWDC 2026 features were excluded.[31][22][23]

Comparison with Galaxy AI and Pixel AI

Apple Intelligence entered a competitive landscape where Google and Samsung had established AI offerings. The comparison below reflects the state of each platform as of mid-2026.

Feature areaApple IntelligenceGoogle Pixel AI (Gemini)Samsung Galaxy AI
Writing assistanceWriting Tools (Rewrite, Proofread, Summarize)Gemini in Google Docs, Gmail Smart ComposeChat Assist, tone adjustment in Samsung Keyboard
Image generationImage Playground, Genmoji, Image WandImagen integration in Google Photos and MessagesAI-generated wallpapers, stickers
Photo editingClean Up (object removal)Magic Eraser, Magic Editor in Google PhotosPhoto Assist (move, resize, delete, and generate fill)
Visual searchVisual Intelligence (now part of Siri AI)Google Lens, Circle to SearchCircle to Search (Google-powered)
TranslationLive Translation in calls, Messages, FaceTimeGoogle Translate integration, live caption translationLive Translate for calls and text
Voice assistantSiri AI (AFM 3 models on device and on Private Cloud Compute)Gemini as default assistantBixby with Galaxy AI, Google Gemini available
Call screeningCall Screening (iOS 26, September 2025)Call Screen (available since 2018 on Pixel)Bixby Text Call (Bixby answers and shows the caller's speech as text)[35]
Privacy modelOn-device processing, Private Cloud ComputeMix of on-device and cloud processing under Google's privacy policiesMix of on-device and cloud processing, with an optional on-device-only setting[36]

Expanded article table

Samsung's Galaxy AI, which launched with the Galaxy S24 in January 2024, gained an early lead in photo editing and real-time call translation. Google's Circle to Search became popular across Samsung and Pixel devices. Apple's competitive advantage centered on its privacy architecture and the Foundation Models framework's zero-cost on-device developer access, though Google's Call Screen had a multi-year head start on Apple's call screening feature.[12]

Reception

Positive reception

Some Apple Intelligence features received strongly positive reviews, particularly from users who found Writing Tools, Smart Reply, and Clean Up genuinely useful in daily use. The Clean Up tool in Photos was widely praised as one of the more reliable object-removal implementations in a smartphone camera app. Writing Tools drew favorable comparisons to Microsoft 365 Copilot for its integration depth, working across any app where text can be edited rather than being confined to first-party software.

The Foundation Models framework was well received by developers, who noted that zero-cost on-device inference opened up AI-powered features for apps that could not previously justify cloud API costs.

Mixed and negative reception

The dominant criticism of Apple Intelligence in 2024 and 2025 was that Apple marketed features aggressively before they existed. The iPhone 16's marketing emphasized Apple Intelligence capabilities, yet many of those capabilities were absent at launch and took months to arrive. The SellCell survey of more than 2,000 smartphone users (including over 1,000 iPhone owners with AI-capable models), fielded November 28 to December 6, 2024, found that 73% of Apple Intelligence users said the features were "not very valuable" or "add little to no value," and that only 41.6% had even tried the features. The same survey found 87% of Galaxy AI users gave similarly low ratings.[2]

The notification summary errors in December 2024 were the most prominent negative story. Summaries attributed false statements to BBC News, The New York Times, and other outlets, including a false claim that a darts player had won a championship before it was played and an incorrect summary stating a tennis player had "come out as gay." The BBC filed a formal complaint. Reporters Without Borders expressed concern about risks to media outlets. Apple disabled summaries for News and Entertainment apps in January 2025.[8]

Reviewers and users also noted that many Apple Intelligence features felt like table stakes compared to what Google and Samsung had already shipped. Call Screening, for instance, arrived in iOS 26 in September 2025, nearly seven years after Google introduced an equivalent feature on Pixel phones.

The Siri delays damaged Apple's credibility specifically. Apple's marketing for the iPhone 16 had placed next-generation Siri at the center of the pitch. When those features slipped to 2026, and then slipped again, the gap between the promised product and the shipped product became a running story in the tech press. Apple settled the resulting class action for $250 million without admitting wrongdoing; the settlement was not a court finding that the advertising was false.[5][29]

Limitations

Apple Intelligence has several technical and functional limitations as of mid-2026:

  • Hardware requirement: The 8 GB RAM minimum excludes iPhone 15 and 15 Plus, iPad mini 6th generation, and Intel-based Macs (every Apple silicon Mac, including 8 GB M1 models, qualifies).
  • Language coverage: While the March 2025 expansion added eight languages, many languages remain unsupported. EU language support was delayed by the Digital Markets Act situation.
  • Internet dependency for some features: On-screen awareness, personal context, and the most capable cloud reasoning require either PCC connectivity or cloud model access. Fully offline use is limited to the features running on the local on-device model.
  • Notification summary accuracy: Summaries for News and Entertainment apps, paused in iOS 18.3, returned in iOS 26 with a warning that summarization may change the meaning of headlines.[38]
  • Genmoji quality: Users and reviewers noted that Genmoji output quality lags behind third-party image generation tools like DALL-E or Midjourney, particularly for photo-realistic styles.
  • China availability: As of mid-2026, Apple Intelligence remained unavailable in China; Apple said in October 2026 that iOS 27.3 would bring some earlier features to iPhones there in December 2026, without Siri AI.[31]
  • EU Siri AI delay: Siri AI on iOS 27 and iPadOS 27 was withheld from the EU at launch over the Digital Markets Act.

Continue reading

More from See also

More to read

Topics:AI Tools & ProductsArtificial Intelligence

References

  1. ^1 ^2 ^3Introducing Apple Intelligence for iPhone, iPad, and Mac - Apple Newsroom
  2. ^1 ^2 ^3iPhone vs. Samsung AI Survey: 73% of iPhone Users Say AI Features Add Little or No Value - SellCell
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Apple Intelligence gets even more powerful with new capabilities across Apple devices - Apple Newsroom
  4. ^WWDC 2026 preview: iOS 27, Gemini-powered Siri and everything else to expect - Tom's Guide
  5. ^1 ^2 ^3 ^4Apple to pay $250M to settle lawsuit over Siri's delayed AI features - TechCrunch
  6. ^1 ^2 ^3 ^4Apple Intelligence is available today on iPhone, iPad, and Mac - Apple Newsroom
  7. ^1 ^2Apple Intelligence: What to Know About Summaries and Smart Replies - MacRumors
  8. ^1 ^2Apple disables AI notifications for news in its beta iPhone software - CNBC
  9. ^1 ^2 ^3 ^4Apple Intelligence now features Image Playground, Genmoji, and more - Apple Newsroom
  10. ^Use Apple Intelligence in Photos on iPhone - Apple Support
  11. ^Use visual intelligence on iPhone - Apple Support
  12. ^1 ^2I'm stoked about iOS 26 Call Screening, but it proves Apple is still playing catch up - Tom's Guide
  13. ^1 ^2 ^3 ^4 ^5Apple delays Siri AI improvements to 2026 - CNBC
  14. ^1 ^2Apple picks Google's Gemini to run AI-powered Siri coming this year - CNBC
  15. ^1 ^2Google wins in AI deal that highlights Apple's own AI struggles - Fortune
  16. ^1 ^2iOS 27 Will Let You Pick Claude or Gemini Instead of ChatGPT for Apple Intelligence - MacRumors
  17. ^1 ^2 ^3 ^4Private Cloud Compute: A new frontier for AI privacy in the cloud - Apple Security Research
  18. ^1 ^2 ^3 ^4 ^5Introducing Apple's On-Device and Server Foundation Models - Apple Machine Learning Research
  19. ^Apple's Foundation Models framework unlocks new intelligent app experiences - Apple Newsroom
  20. ^Apple Explains iPhone 15 Pro Requirement for Apple Intelligence - 3uTools
  21. ^Apple Intelligence won't launch in EU in 2024 due to antitrust regulation - CNBC
  22. ^Apple works with Alibaba to bring Apple Intelligence to China - AppleInsider
  23. ^Apple Intelligence finally rolls out in China - AppleInsider
  24. Comparing AI in smartphones: Galaxy AI vs Google AI vs Apple Intelligence - musicMagpie
  25. Apple Intelligence Continues To Evolve, Though Its Rollout Has Faced Delays - SimplyMac
  26. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Everything announced at Apple's WWDC 2026 keynote - Engadget
  27. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Introducing the Third Generation of Apple's Foundation Models - Apple Machine Learning Research
  28. ^1 ^2Apple's new foundation models don't contain a drop of Gemini, as we said they wouldn't - AppleInsider
  29. ^1 ^2iPhone users could get up to $95 per device as Apple reaches $250M settlement over Siri delays - 9to5Mac
  30. ^1 ^2 ^3 ^4Apple Releases iOS 27 and iPadOS 27 With Siri AI and Liquid Glass Update - MacRumors
  31. ^1 ^2Apple Says iOS 27.3 Beta Coming Next Week - MacRumors
  32. ^1 ^2 ^3Apple's New AI Models Contain 'None' of Google's Gemini Assistant - MacRumors
  33. ^1 ^2Updates to Apple's On-Device and Server Foundation Language Models - Apple Machine Learning Research
  34. ^1 ^2Gurman: Apple's new Siri delayed again, will not be included in iOS 26.4 - GSMArena
  35. ^Use Bixby text call on your Galaxy phone - Samsung US Support
  36. ^Samsung Galaxy S24's AI features can be set to on-device-only processing - Tom's Guide
  37. ^Apple to Reportedly Pay Google $1 Billion a Year for Siri's Custom Gemini AI Model - APH Networks
  38. ^1 ^2Apple brings back notification summaries for news in iOS 26 - AppleInsider

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

9 revisions · v9 · 7,289 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check of the full article against its cited sources (170 claims, 44 sources fetched); 12 material + 34 minor fixed

Cite this page: AI Wiki. "Apple Intelligence." aiwiki.ai, updated 28 Oct 2026, fact-checked 2 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/apple_intelligence

Edit article