From custom model to foundation model
For years, Core ML was Apple's official gateway to on-device inference on Apple Silicon — a layer designed for developer-trained models imported in `.mlmodel` format. With iOS 27 and macOS Golden Gate, Apple announced Core AI as its direct replacement, rebuilt around the reality of large foundation models — what Apple calls Apple Foundation Models (AFM).
As reported by 9to5Mac ahead of the keynote and confirmed during technical sessions, the new framework exposes APIs more consistent with long-context handling, multi-step reasoning, and dynamic routing between on-device execution and Private Cloud Compute. This is not a simple rebrand: the internal architecture has been revised to align with the `LanguageModel` protocol introduced in Swift and already present in the Foundation Models APIs open to developers.
What changes for developers
Developers who used Core ML for classic tasks — image classification, regression, object detection — will find a documented migration path from Apple, but the message is clear: for anything involving language, text generation, or multimodal understanding, Core AI is the starting point. The Xcode 27 tooling ecosystem is already updated accordingly, with the Coding Assistant able to suggest Core AI integration where Core ML would previously have been used.
The change also has performance implications: Core AI is optimized to leverage the new Neural Engines in A18, A19, and M4/M5 chips, reducing inference latency compared to the old stack.