Apple expands on-device generative model support for third-party apps in iOS SDK

AI · 5 min read

At WWDC-style developer sessions this week, Apple introduced a new On-Device Models Framework in the iOS SDK that standardizes packaging, versioning, and secure execution of small generative models across iPhone and iPad. The framework includes runtime guarantees, size thresholds, and a signed manifest to prevent misuse while enabling offline experiences.

Apple demonstrated the framework with examples for image retouching, voice synthesis for accessibility, and local text generation inside messaging apps. Developers can now choose between model families optimized for performance or battery preservation, and the runtime will transparently switch between hardware (Neural Engine) and CPU when necessary.

Privacy is central: models remain sandboxed inside the app container and cannot access system-level telemetry. Apple also announced an App Store guideline extension requiring that any shipped model include a clear developer-supplied description of training data categories and permitted use cases to help reviewers and users understand the AI’s provenance and limitations.