macOS Sequoia adds native Swift-ML runtime and sandboxed model store

AI · 5 min read

macOS Sequoia includes a new Swift-ML runtime optimized for Apple silicon that runs models with lower overhead and native Swift bindings. The runtime targets common tasks like on-device vision, speech, and small LLMs, and integrates tightly with Core ML for conversion and optimization pipelines.

A sandboxed system model store lets apps request curated system models at runtime without bundling large artifacts, reducing app size. Models are signed and versioned by Apple, and apps can request permissions to cache models locally; sandboxing prevents cross-app model access to reduce privacy and intellectual property risk.

Developers will find faster prototyping with prebuilt Swift playground templates, and Apple’s documentation includes migration guidance from Python/C++ inference stacks. The focus is clearly on encouraging native app ML that stays on-device for privacy and performance.