Core AI (iOS 27+)

Core AI is a memory-safe Swift framework for loading, specializing, and running your own models fully on device. Requires iOS 27 / macOS 27 and Xcode 27. Core ML stays supported for existing models; evaluate Core AI for new LLM-style or custom-compute workloads.

Contents

• Types
• Load and run
• Neural Engine behavior in iOS 27
• Debugging
• Common mistakes
• Sources

Types

Load and run

1. Create an AIModel from a file URL with init(contentsOf:options:), or call specialize(contentsOf:options:cache:cachePolicy:) to reuse a cached specialization.
2. Get a function with loadFunction(named:); list names with functionNames.
3. Inspect inputs and outputs with functionDescriptor(for:) or InferenceFunction.descriptor.
4. Call run(inputs:states:outputViews:) with NDArray inputs. The call is async.

Specialize ahead of first use and cache the result; specialization is the slow step. To drive an LLM through the Foundation Models session API instead, see foundation-models-ios27.md.

Neural Engine behavior in iOS 27

From the iOS 27 release notes:

• Background Neural Engine access is restricted, like GPU limits. Background inference needs the entitlement com.apple.developer.background-tasks.continued-processing.inference.
• Loading models over 1 GB is faster.
• Neural Engine memory counts against your app process and shows in the Allocations instrument.

Debugging

Core AI Debugger, Instruments, and a debug gauge are available in Xcode 27. Use the Foundation Models instrument when the model runs inside a LanguageModelSession.

Common mistakes

• Starting Neural Engine work from a background task without the inference entitlement.
• Specializing on every launch instead of using AIModelCache.
• Budgeting memory as if Neural Engine allocations were free; they are attributed to the app.
• Rewriting working Core ML pipelines without a measured reason.

Sources

• https://sosumi.ai/documentation/coreai
• https://sosumi.ai/documentation/ios-ipados-release-notes/ios-ipados-27-release-notes