Apple announces developer access to its on-device model

On June 9, 2025, Apple announced Foundation Models, a framework for apps to access the on-device Apple Intelligence model. The original source supports the documentary core of the event; the announcement described a capability of forthcoming system releases, not immediate universal availability.

Until now, that on-device model — the one powering much of Apple Intelligence — was Apple's exclusive territory. Starting today, according to the company's announcement, any app will be able to call it with just three lines of Swift code. No cost, no need for an internet connection, and no user data ever leaving the device.

Why this is different from calling a chatbot's API

Most apps that integrate generative AI today do so by calling an external server — OpenAI, Anthropic, or whichever provider. That means a cost per query, dependence on an internet connection, and user data leaving the device.

Apple's approach is the opposite. The model lives on the phone itself, so it responds even without coverage, generates no API bill, and — because everything is processed locally — the company says it protects the privacy of users' conversations and content. Craig Federighi, Apple's senior vice president of Software Engineering, put it this way in the announcement: "We're also taking the huge step of giving developers direct access to the on-device foundation model powering Apple Intelligence, allowing them to tap into intelligence that is powerful, fast, built with privacy, and available even when users are offline." According to Federighi, this "will ignite a whole new wave of intelligent experiences" in the apps people use every day.

The framework and the rest of the Apple Intelligence updates are available for testing starting today, and will reach users with compatible devices and languages this fall, Apple said.

The rest of the package: translation, Genmoji and visual intelligence

The Foundation Models framework didn't arrive alone. Apple also introduced Live Translation, a real-time translation system built into Messages, FaceTime and Phone, powered by Apple's own models that run entirely on the device. In Messages, it automatically translates as you type; in FaceTime, it shows translated captions while you still hear the original voice; on Phone calls, the translation is spoken aloud during the conversation.

Genmoji and Image Playground, the tools for creating custom images and emoji, are getting an expansion: users can now mix emoji together and combine them with text descriptions, plus new styles — like oil painting or vector art — available through ChatGPT. Apple stresses that nothing is shared with ChatGPT without the user's permission.

Visual intelligence, which already let users identify objects and places through the camera, now extends to whatever appears on the iPhone screen: users can ask ChatGPT about something they're looking at, search for similar products on Google or Etsy, or let the system detect an event in an image and suggest adding it to Calendar, automatically pulling out the date, time and location.

On Apple Watch, Workout Buddy makes its debut — an assistant that combines data from an ongoing workout with a user's fitness history to generate real-time motivational commentary, using a synthetic voice built from recordings of Fitness+ trainers. It will launch in English for activities like running, walking, outdoor cycling, HIIT and strength training, and requires Bluetooth headphones and a compatible iPhone nearby.

Apple Intelligence will also add eight more languages by the end of the year: Danish, Dutch, Norwegian, Portuguese (Portugal), Swedish, Turkish, Traditional Chinese and Vietnamese.

What changes for app developers

Opening up the on-device model is the piece with the biggest medium-term implications. Hundreds of millions of iPhones, iPads and Macs already have a language model installed that can summarize text, generate content or reason about information — and until today, only Apple could use it inside its own apps. By opening it to third parties at no API cost, any developer can add generative AI features to their app without negotiating contracts with model providers or worrying about the monthly bill that usually comes with that kind of integration.

The trade-off is that the model lives within Apple's ecosystem: it only works on Apple devices and under Apple's terms. For a developer already building for iPhone, it's a free way to add intelligence without relying on third parties. For the market of API model providers, it's a competitor that doesn't charge per query.

Turning the headline into a check

The first step is to freeze the system's identity. Apple announced Foundation Models, a framework for apps to access the on-device Apple Intelligence model. A commercial name may cover different revisions, automatic routes and tools. A test record should preserve date, access mode, configuration, permissions and the full output. Without that snapshot, an improvement or failure observed today cannot rigorously be attributed to the version another person will use tomorrow.

Next, turn how to check device, operating system, privacy, limits and fallback before adopting a local API into cases with acceptance criteria. Build a local sample containing easy, ambiguous, long and deliberately impossible tasks. Record the input, the information the model may consult and what outcome would count as sufficient. A vendor-selected demonstration shows possibility; a test set preserved by the user measures reliability.

Autonomy needs a permission ladder. Reading and proposing are not the same as editing, sending or buying. A safer setup begins with read-only access, requires a preview and reserves execution for explicit approval. It also keeps a log and a rollback path. Judge the model by the errors the surrounding system contains, not by the confidence of its plan.

What the record must preserve

Cost and quality must be measured together. A cheap answer that must be reviewed from scratch can cost more than a slower but verifiable one. Measurement includes waiting, retries, consumption, human oversight and the consequences of failure. the announcement described a capability of forthcoming system releases, not immediate universal availability. That boundary turns the announcement into a testable hypothesis rather than a promise to be believed.

An evidence sheet separates four columns: what the source claims, what it shows, what it did not measure and what would change the conclusion. That discipline prevents an absence from becoming a promise and a condition from vanishing in summary. It also lets the story be updated without rewriting history from a later outcome.

Include a negative case before deciding. Find a situation where the system, rule, transaction or study does not meet the need and record the signal that would require stopping. Selected successes show that something can happen; the negative case reveals the boundary and lowers the cost of discovering it after deployment.

The skill that outlasts the announcement

A valid comparison preserves denominator and axis. It does not pit a point figure against an average, future capacity against installed capacity or a forecast against an observation. When two sources use similar language, reconstruct what they counted and over what period. If those differ, publish them as different measures instead of inventing a ranking.

The record should survive a version change. Keep URL, consultation date, document, configuration and decision. When new evidence appears, add it with its date and explain what it changes. That traceability prevents opposite errors: keeping an expired conclusion or pretending later information was known on the event date.

The transferable skill in this story is how to check device, operating system, privacy, limits and fallback before adopting a local API. The procedure is short: name the document, preserve the date, fix the axis, find the condition and design a check that can fail. With those steps, a reader need not accept or reject the announcement by intuition; the decision follows a visible chain of evidence.

Before closing, another person should be able to reconstruct the conclusion without knowing the headline. Give them the sources, conditions and negative case, then ask what they would accept and reject. If they need an assumed intent, a figure without a denominator or an undated later fact, the chain still has a gap. That short review catches errors that fluent prose can conceal.