Apple WWDC 2025: Full Breakdown of On-Device AI Models and the Foundation Models Framework

Published June 9, 2025byMergerank AI
AI & Search
Apple WWDC 2025: Full Breakdown of On-Device AI Models and the Foundation Models Framework

At Apple’s WWDC 2025 conference, the company detailed its shift toward integrating generative AI directly on devices. One of the key announcements was the introduction of the Foundation Models framework, a Swift-based developer interface that grants access to Apple’s own large language models (LLMs) for use in third-party apps. This marks a significant technical change in how AI can be implemented across Apple platforms—especially in privacy-sensitive contexts.

What Is Apple’s Foundation Models Framework?

The Foundation Models framework is a new set of APIs designed for use within Apple’s ecosystem (iOS, iPadOS, macOS, and watchOS). Developers can access Apple’s proprietary LLMs directly on-device, without sending data to the cloud. This design reduces dependency on internet connectivity and mitigates risks related to data privacy and third-party servers.

The framework is written in Swift and is designed to integrate easily into existing app architectures. Developers can call LLM-based features such as text summarization, question answering, context-aware automation, and image generation—all locally on supported Apple hardware.

How Do Apple’s On-Device AI Models Work?

According to Apple, the models accessible through the framework are derived from a compact ~3 billion parameter foundation model. First detailed in a 2024 technical report, this model is optimized for mobile inference. It supports:

  • Text generation and natural language understanding
  • Summarization and content transformation
  • Tool invocation within apps
  • Image creation and manipulation

All inference happens on the user’s device, improving response time, avoiding API costs, and aligning with Apple’s privacy-first design.

Use Cases for Developers

Apple highlighted several specific scenarios during the event to demonstrate potential use cases for developers:

  • Education apps generating quizzes from student notes.
  • Outdoor apps offering trail suggestions via natural language queries.
  • Writing tools that provide real-time rephrasing or summarization.
  • Messaging and call apps with built-in translation features.

Integration into System Features

These same models are used internally to power features such as:

  • Visual Intelligence for recognizing and describing images.
  • Writing Tools for editing and refining text.
  • Shortcuts that respond to natural language.
  • Genmoji creation from descriptive prompts.

Apple Watch Application: Workout Buddy

In watchOS 26, the Workout Buddy feature uses local AI to analyze fitness history and offer tailored recommendations—all without needing a network connection.

Frequently Asked Questions (FAQ)

Q: What is Apple’s Foundation Models framework?

A: It’s a developer API introduced at WWDC 2025 that allows apps to access Apple’s on-device large language models for text generation, summarization, and more—all without cloud processing.

Q: Are Apple’s AI models cloud-based?

A: No. Apple’s models are designed for local execution, meaning they run entirely on-device. This preserves privacy and ensures the models function without an internet connection.

Q: How big is Apple’s on-device model?

A: The model is approximately 3 billion parameters, making it compact enough for mobile use but still capable of complex tasks like summarization and tool calling.

Q: Can third-party developers use the same AI as Apple’s apps?

A: Yes. Apple has made the same model used in system features like Genmoji and Shortcuts available to developers through the Foundation Models framework.

Q: What kinds of apps can use on-device AI?

A: Any app that benefits from language understanding or generation—such as education, health, productivity, or communication apps—can use the framework to add intelligent features.

Q: Does the use of this model require internet access?

A: No. The AI functions run offline, providing speed, cost savings, and user privacy.