Apple opens on-device AI model to developers and adds real-time translation

Apple Intelligence is taking off the training wheels, and this time developers get to drive.

At WWDC 2025 on June 9, Apple rolled out a substantial set of updates to its AI ecosystem across iPhone, iPad, Mac, and Apple Watch. Beyond flashy new tricks like on-screen search, system-wide Live Translation, and workout-coaching audio, the headline event is clear: Apple is officially opening up its core on-device foundation model to third-party app developers through a native Swift framework.

Screen search and real-time translation

For everyday iPhone users, visual intelligence is finally stepping beyond the physical camera viewfinder. By pressing the standard screenshot buttons, users can now trigger visual intelligence to scan whatever is currently on display. You can circle a lamp to find it on Google or Etsy, ask ChatGPT to explain a diagram in your feed, or let iOS parse dates and locations from an event flyer directly into your Calendar app.

Communication is getting an on-device overhaul too. Live Translation is moving natively into Messages, FaceTime, and the Phone app. In Messages, your texts translate as you type, and incoming replies are instantly converted back. FaceTime adds translated live captions, while regular phone calls now feature real-time spoken translations. Because these models run completely on-device, conversations remain private.

Other fun updates include emoji mashups in Genmoji—like merging a sloth with a lightbulb—and new artistic styles in Image Playground powered by ChatGPT. On the wrist, Apple Watch is gaining Workout Buddy, which pairs with Bluetooth headphones and uses generative voice models trained on Fitness+ trainers to deliver real-time, personalized motivation based on your pace and heart rate.

Three lines of code and zero API bills

While consumers get neat party tricks, developers get the real payload. Apple introduced the Foundation Models framework, granting apps direct access to the on-device large language model powering Apple Intelligence.

Apple claims developers can tap into the model using as few as three lines of Swift code. The framework natively supports guided generation and tool calling. Crucially, inference runs locally on the Apple Silicon chip, meaning it works entirely offline, protects user privacy, and costs developers zero cloud API fees.

Shortcuts also received a deep upgrade, allowing power users to feed Apple Intelligence models—running either locally or via Private Cloud Compute—into automated workflows. In addition, Apple confirmed that Apple Intelligence will expand to eight more languages by year-end, including Traditional Chinese, Vietnamese, Danish, Dutch, Norwegian, Portuguese, Swedish, and Turkish.

Why it matters

We think this is the smartest move Apple could have made after a bumpy first year with AI. A year after its debut, Apple Intelligence had largely struggled to meet high expectations: the marquee overhaul for Siri was delayed past March, and early features like notification summaries felt unpolished. Instead of doubling down on trying to out-reason frontier cloud models, Apple leaned straight into its greatest competitive moat: hardware silicon and developer lock-in.

For app builders, free on-device inference is a game-changer. Small development teams and bootstrapped indie makers can now weave summarization, extraction, and local natural-language search into their apps without racking up massive monthly OpenAI API bills. That alone could spark a creative wave of smart iOS utilities that simply cannot exist economically on web-only architectures.

The catch? Local on-device models inherently have a performance ceiling. Whenever complex, nuanced reasoning is required, users will still find themselves bouncing to ChatGPT or waiting on Apple's delayed Siri promises. But by turning every iPhone into an offline, zero-marginal-cost AI workstation for developers, Apple just bought itself a huge operational head start over competitors trying to monetize every single cloud token.