iOS 20 Goes On‑Device
Apple’s iOS 20 pushes generative and language tasks onto the device — reviewers report native APIs for image generation, summaries and simple app logic to cut latency and protect privacy. The OS is being positioned as a privacy‑first push toward local NLP and vision workloads for developers. (youtube.com)
Apple exposed on‑device large language model access through a new FoundationModels framework at WWDC 2025, and the framework shipped as part of the OS update Apple distributed under iOS 26. The on‑device language model Apple documents weighs in at roughly 3 billion parameters, with Apple maintaining a larger server model for more complex tasks via Private Cloud Compute. (machinelearning.apple.com) Image generation is available to apps through the ImagePlayground framework and the ImageCreator API, which let developers present a system image sheet or programmatically return generated images to their Swift code. Language features exposed to developers include summarization, extraction and “guided generation,” plus tool‑calling and streaming “snapshot” responses that can produce typed Swift values via @Generable annotations. (developer.apple.com; wwdcnotes.com) Apple positions the stack to run locally for low latency and offline operation, and the company limits Apple Intelligence features to “Apple Intelligence‑compatible” devices rather than every iPhone that can run the OS; compatibility lists published with iOS 26 show a narrower set of models for full AI features. Apple highlighted third‑party integrations—apps such as SmartGym, Stoic and VLLO have announced features built on FoundationModels—and community tooling (including a React Native preview package) surfaced within months of the WWDC reveal.
