Apple Developer Resources

The definitive reference for Core ML. Covers model loading, MLModel API, MLModelConfiguration for compute unit targeting, and performance profiling. The starting point for any on-device inference work.

Swift API for Apple's on-device language model (iOS 26+). Covers the Generable protocol for structured generation, session management, and the LanguageModelSession API. The correct choice for generative text tasks — no model management required.

The Python toolkit for converting PyTorch, TensorFlow, and ONNX models to Core ML format. The ct.optimize namespace handles quantization, palettization, and pruning. Use this to prepare any custom model for on-device deployment.

The core class you instantiate for inference. Critical reading: MLModelConfiguration (sets computeUnits to target the Neural Engine), and the async prediction API added in iOS 16 that eliminates the need to manually dispatch to a background thread.

Apple's high-level computer vision API. When your Core ML model processes images, use Vision rather than calling Core ML directly — Vision handles image preprocessing, orientation correction, and coordinates the pipeline automatically.

Apple's on-device NLP framework. Before writing a custom NLP Core ML model, check whether NLTagger, NLEmbedding, or NLModel already does what you need. The embedding APIs are particularly useful for semantic search.

The best single resource on Core ML optimization. Covers FP16 conversion, palettization, and pruning with concrete performance numbers. The benchmark showing 4× size reduction with less than 2% accuracy loss comes from this session.

Covers the async prediction API, compute unit configuration, and performance profiling with the Core ML Performance Report in Xcode. Essential for anyone moving beyond tutorial-level Core ML usage.

Overview of the full ML stack as of 2024: Core ML, Create ML, Vision, Natural Language, and the trajectory toward Foundation Models. Useful for understanding how the layers fit together before deciding where to build.

The predecessor session to the 2023 version. Covers the mlprogram format introduction and the shift away from neural_network format. If you maintain older models, this explains why the migration to mlprogram matters.

Introduces the Neural Engine performance tools and the A15 Bionic ML capabilities. The benchmark baselines for ANE inference (35 TOPS on A15) originate from this era of Apple's public specifications.

The canonical reference for async/await, actors, and structured concurrency. The actor isolation model is the correct pattern for wrapping Core ML's non-thread-safe MLModel — this document explains why actors solve the problem cleanly.

The Swift type for bridging callback-based APIs to async/await. Used in on-device LLM integration (bridging llama.cpp's C token callbacks to SwiftUI) and any other producer/consumer pattern where values arrive over time.

The original proposal that introduced actors to Swift. Reading the Motivation and Proposed Solution sections clarifies the actor isolation guarantees that make Core ML thread-safety problems disappear when using this pattern.

The foundational async/await proposal. Understanding the execution semantics — how continuations are scheduled, when a suspension point is encountered — is essential for reasoning about inference performance and thread behavior.

The definitive SwiftUI reference. The .task modifier and @Observable macro (added in SwiftUI 5 / iOS 17) are the current standard for connecting async data sources — including AI inference streams — to SwiftUI views.

Apple's modern data persistence framework (iOS 17+). The correct layer for storing inference results — AI-generated categories, scores, and embeddings — as first-class entities that compose with SwiftUI's @Query macro.

The authoritative reference for what Apple accepts and rejects. Section 4.2 (Minimum Functionality) and Section 5.1 (Privacy) are the most commonly cited in review decisions. Required reading before any first submission.

Apple's guide to code signing certificates, provisioning profiles, and the relationship between them. The source of truth for signing identity requirements per distribution method (Debug, TestFlight, App Store, Direct Distribution).

Required reading for macOS tools distributed outside the App Store. Covers the Hardened Runtime entitlement, notarytool CLI workflow, stapling, and Gatekeeper behavior. A notarization failure silently breaks distribution for most users.

Documents the runtime protection policies enforced by the Hardened Runtime entitlement. Certain APIs (JIT compilation, unsigned executable memory) require explicit entitlement exceptions. Relevant when integrating native libraries like llama.cpp.

Complete reference for all entitlements: HealthKit, iCloud, Push Notifications, App Groups, and more. Entitlement mismatches between capability configuration and the .entitlements plist file are a leading cause of App Store rejection.

The persistence layer used in production AI-native apps for storing inference results alongside source data. NSFetchRequest with sort descriptors on AI-generated fields — relevance scores, predicted categories — is the foundation of AI-powered list ordering.

The CloudKit-backed Core Data container. Relevant for the decision of whether to sync AI inference results or recompute locally — syncing results is faster on new devices but increases CloudKit storage and requires conflict handling.

Apple's cloud database for iOS and macOS. For apps with server-side data needs beyond iCloud sync (shared data between users, public databases), this is the correct Apple-first path before considering third-party back-ends.

The correct API for running ML inference batches in the background (processing new records while the app is not in the foreground). A properly scheduled background task is essential for AI-native apps that process data incrementally.