The most consequential AI announcement at WWDC 2025 was not a single feature but an opening: Apple said third-party apps get direct access to the large language model that powers Apple Intelligence on device. The vehicle is the new Foundation Models framework, and its pitch is unusual — inference at no cost, full offline operation, and data that never leaves the device.
Software engineering chief Craig Federighi called it “the huge step of giving developers direct access to the on-device foundation model,” and predicted it would “ignite a whole new wave of intelligent experiences in the apps users rely on every day.” He also set the premise: the models behind Apple Intelligence are becoming more capable and efficient, and are being woven into more places across each operating system. The framework entered testing the day of the keynote and reaches users this fall on supported devices.
What the Foundation Models framework does
The framework is native to Swift, and Apple says as few as three lines of code can call the model. The WWDC session “Meet the Foundation Models framework” walks through the core capabilities in concrete terms: guided generation that produces output conforming to Swift data structures, snapshot streaming for responsive interfaces as long outputs generate, tool calling for bringing in external data sources, and stateful sessions that manage context across multiple turns. The session listed no prerequisites and came with a code-along and a deep-dive follow-up.
The pattern here is familiar. Structured output and function calling are what developers already use against cloud APIs; Apple lifted those modes onto the device and tied them together with the Swift type system, so parsing and validation shrink to near zero. The migration work is less about the interface and more about repositioning what you expect from the model.
Free inference, no cloud budget
Apple’s announcement leans on one phrase for this model: “powerful, fast, built with privacy, and available even when users are offline.” Inference is explicitly free of cost — there is no per-call cloud API bill. The company’s two examples make the intended use cases clear: an education app generating personalized quizzes from a user’s notes with no cloud API costs, and an outdoors app offering natural-language search that keeps working with no network.
For product teams this changes the cost structure of shipping AI features. Until then the options were a cloud API or a self-compressed, self-maintained local model; now there is a third path hosted by the operating system and updated with it. The trade-off is a hard ceiling on capability — this is a small on-device model suited to high-frequency, lightweight, well-defined tasks, not open-ended long reasoning. Those heavier cases still belong to Private Cloud Compute or third-party cloud APIs.
Apple Intelligence features get their own updates
The user-facing side moved too. Live Translation arrives in Messages, FaceTime, and Phone: messages translate automatically as you type and are delivered in the recipient’s preferred language, with replies instantly translated; FaceTime calls show live translated captions while the speaker’s original voice stays audible; and phone calls get the translation spoken aloud. The whole experience runs on Apple-built models entirely on device, which Apple says is exactly why personal conversations stay personal.
Also in the release: visual intelligence extends to more of what is on your screen, Image Playground and Genmoji gained expressive upgrades, and Shortcuts added intelligent actions — users can call the on-device model or Private Cloud Compute directly and feed the response into the rest of a shortcut, with ready-made actions like summarizing text with Writing Tools or creating images with Image Playground. Apple also committed to bringing Apple Intelligence to eight more languages by the end of the year: Danish, Dutch, Norwegian, Portuguese (Portugal), Swedish, Turkish, Chinese (Traditional), and Vietnamese.
The missing Siri, and how this aged
The honest context: the more personalized Siri did not appear at this keynote, and TechCrunch’s same-day headline said so plainly — “Still no AI-powered, more personalized Siri from Apple at WWDC 25.” Apple shifted the spotlight to what developers could use immediately.
What shipped instead was a bounded, modest on-device model. But with guided generation and tool calling attached, it covers the high-frequency work — summarization, classification, extraction — that most apps actually need. Judged from 2026, the framework did become a standard way for iOS apps to ship AI features, and the offline and privacy arguments kept it credible wherever data cannot leave the device.
Sources
- Apple Newsroom: New Apple Intelligence capabilities
- Apple Developer: Meet the Foundation Models framework
- TechCrunch: Live Translation across Messages, FaceTime, and Phone
- TechCrunch: Still no more personalized Siri at WWDC 25
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.