At the WWDC 2025 keynote, the OS-wide design overhaul "Liquid Glass"—the first in a decade—has been generating major buzz, but perhaps an equally important announcement was made quietly. Today Apple unveiled the "Foundation Models framework," and this is truly a revolutionary announcement: it opens up the on-device AI model at the heart of the personal AI permeating every corner of the iPhone and Mac, "Apple Intelligence," to third-party developers around the world.

This will allow developers to incorporate advanced AI features—fast, ultra-private, and functional even offline—into their own apps with remarkable ease, and without incurring any cloud API costs whatsoever.

This has the potential to fundamentally overturn the conventions of app development and elevate the user experience to a new dimension. And it might be seen as a blow Apple has struck against Google's Android, which has long led the AI space, on the new battleground of "developer experience" and "ecosystem integration."

AD

What Is the "Foundation Models Framework," the True Star of WWDC25?

The "Foundation Models framework" announced by Apple is a new toolset that allows developers to directly access the on-device large language model (LLM) that powers "Apple Intelligence." Advanced AI features that were previously the exclusive domain of Apple's own apps have finally been unlocked for third parties.

For developers, the benefits are immeasurable.

  • Ultimate privacy: Since all processing is completed entirely on the user's device, sensitive data is never sent to an external server. This is the very essence of Apple's privacy philosophy.
  • Overwhelming performance: With no network latency, AI features respond instantly.
  • Full offline support: AI features work without a hitch even in environments without an internet connection. Whether on a plane or in the subway, the app's intelligence remains uncompromised.
  • Zero cost: Developers don't need to pay expensive cloud AI API usage fees. This makes it easier for small-scale developers and individuals to take on the challenge of building apps with advanced AI features.

Traditionally, incorporating generative AI into an app meant calling an external cloud API (such as OpenAI's or Google Cloud's). However, this always came with concerns about network latency, cost, and privacy. Apple has overturned this convention with an "on-device" approach.

Is "Just 3 Lines of Code" True? The Astonishing Ease of Implementation

In the keynote, Apple emphasized just how easy this framework is to use. According to Apple's official documentation and press release, developers can call Apple's on-device LLM with just three lines of Swift code.

This is no mere exaggeration. The Foundation Models framework is deeply and natively integrated with Swift, Apple's programming language. This frees developers from routine tasks like handling complex API requests and data format conversions, letting them focus on the essential question of "what they want to do" with AI.

For example, ideas that were previously difficult or too costly to realize—such as a study app that automatically generates quizzes from a user's notes, or an outdoor app that searches offline for the optimal hiking trail based on natural-language conditions—are now becoming realistic possibilities.

AD

The Heart of Apple's On-Device AI: A Technical Deep Dive

This ease of use is the fruit of the vertical integration of hardware and software that Apple has cultivated over many years. A technical paper published by Apple Machine Learning Research reveals details of what lies at its core.

The models powering Apple Intelligence come in two main types.

  1. On-device model: A model with approximately 3 billion (3B) parameters, optimized for efficiency, that runs directly on Apple Silicon in the iPhone, iPad, and Mac. This is the model that has now been opened up to developers.
  2. Server model: A high-performance, server-based model for handling more complex tasks. This runs on Apple's own secure infrastructure called "Private Cloud Compute," processing requests that the on-device model cannot handle while maximally protecting privacy.

Of particular note is the training process behind these models. Apple has explicitly stated that it does not use any of a user's private data or usage history to train its models. The training data consists only of content from publishers under licensing agreements, publicly available datasets, and public information collected by the company's web crawler, "Applebot." This clearly demonstrates Apple's stance of prioritizing privacy above all else.

Furthermore, the on-device model makes extensive use of quantization techniques to minimize memory usage while maintaining performance. The decoder weights are compressed to an astonishing average of 2 bits per weight, making it possible to run a large-scale model smoothly on the limited resources of a smartphone.

Two Revolutionary Features That Change the Rules of Development: "Guided Generation" and "Tool Calling"

The true value of the Foundation Models framework goes beyond mere ease of implementation. It comes with two powerful features that dramatically boost developer productivity and expand the possibilities of AI.

Type-Safe AI Output: The Impact of "Guided Generation"

One of the biggest challenges in conventional generative AI development has been controlling the format of AI output. Even when instructing an AI to output in JSON format, the format would sometimes break down or return unexpected data, forcing developers to spend enormous effort on parsing and error handling.

Apple's "Guided Generation" elegantly solves this problem. Developers simply define the data structure (a struct or enum) they want as output in their Swift code and mark it with @Generable.

// Example: Define a quiz question, its options, and the correct answer
@Generable
struct QuizItem {
    var question: String
    var options: [String]
    var answer: String
}

With just this, the framework works behind the scenes to prompt the LLM to generate output conforming to this QuizItem structure, guaranteeing that the generated result always conforms to this type. Developers are completely freed from tedious parsing work and can receive the AI's output directly as a type-safe Swift object. This is a truly revolutionary feature that dramatically improves both development efficiency and application stability.

Infinitely Extending the Model's Capabilities: The Possibilities of "Tool Calling"

Another powerful feature is "Tool Calling." This is a mechanism for giving a general-purpose LLM app-specific "tools" to extend its capabilities.

Developers can define their own tools by writing code that conforms to the Tool protocol. For example, consider the following kinds of tools.

  • A tool that accesses an internal database to check product inventory
  • A tool that hits a web API to fetch the latest weather forecast
  • A tool that executes a specific function within the app (e.g., playing music, editing a photo)

When a user requests, "What's the weather in Tokyo today? Also, play some relaxing music," the LLM interprets this and decides to call the "weather forecast tool" and the "music playback tool" in the appropriate order. The framework automatically optimizes and manages these complex tool-calling sequences (whether they should run in parallel or in series).

This makes it possible for a general-purpose on-device LLM to behave as intelligently as if it were a "dedicated AI" built solely for that app. Apps like the journaling app "Day One" and the hiking app "AllTrails" are already leveraging this technology and beginning to offer more intelligent experiences.

AD

Why Does Apple Insist on "On-Device"? Is This a Challenge to Android?

Behind Apple's steadfast commitment to on-device AI lies both the company's unwavering philosophy and a clear strategy against its competitors. The benefits of privacy, performance, and zero cost have already been mentioned, but this approach stands in stark contrast to the cloud-centric AI strategy led by Google.

Google boasts one of the world's top-performing AI models, "Gemini," and has provided its power through cloud APIs. It has also been working on on-device AI processing on Android devices (Gemini Nano). However, in terms of "developer experience"—how easily and seamlessly third-party developers can integrate that power into their apps—Apple may have gotten a step ahead.

The ease of implementation with "just 3 lines of code," the guarantee of type-safe output through "Guided Generation," and the deep ecosystem integration enabled by "Tool Calling"—these hold extremely practical and powerful appeal for engineers caught up in day-to-day development work. As Android Authority put it, "Apple has overtaken Android"—Apple isn't competing merely on the raw performance of AI models, but has launched a higher-level battle over "how easily developers can deliver a superior AI experience to users."

AI App Development Enters Its "Second Chapter." The Future Apple Envisions

The arrival of the Foundation Models framework signals that AI app development has entered a new era—its "second chapter." AI is no longer a special technology that only a handful of experts can handle. An era has begun in which every developer can incorporate intelligence into their apps as a matter of course, as naturally as placing a UI component.

Apple has generously handed the "key to AI" to its partners—developers—to further enrich its own ecosystem. This decision will make the appeal of Apple's platform unparalleled, drawing users even more strongly into that ecosystem.

From here on, we're bound to witness a wave of innovative applications, beyond anything we could have imagined, created by developers now wielding this new power. What kind of surprises will your favorite app on your smartphone bring you next, thanks to this technology?

Sources