On-Device Intelligence – shipping local AI on Apple platforms, compiler-verified

On-Device Intelligence: The Apple Developer's Guide to Shipping Private, Local AI

Why this book

The framework docs tell you what compiles. This book tells you what ships.

In 2026 every iOS team faces the same decision: Apple's free on-device model, an open model you control, or a cloud API with a per-request invoice. The difference between a demo and a product is a hundred unglamorous details: jetsam budgets, context-window overflow, availability ladders, model swaps arriving silently in OS updates.

Every code snippet is compiler-verified against the shipping SDK, including the places where Apple's own sample code and popular blog posts no longer compile.

Real numbers, stated method: benchmark tables measured by an open harness you can re-run, not leaderboard folklore.

The whole lifecycle: architecture decision, implementation, memory and thermals, privacy claims that survive legal review, App Review notes, and regression evals with no cloud in the loop.

Free updates: the iOS 27 revision (the LanguageModel protocol, Core AI) ships to every buyer at no charge.

Contents

Twelve chapters, four parts, three runnable tools.

Part I · The Decision

Why Local: the economics, the latency floor, the true capability ceiling

The Apple AI Stack, Mapped: one decision matrix for FM, MLX, Core ML, llama.cpp

Your First On-Device Feature FREE

Part II · Foundation Models in Production

Guided Generation Deep Dive

Tool Calling

Sessions, Context, and Memory

Availability and the Device Matrix

Part III · Owning the Model

MLX Swift: Bring Your Own Model

Ollama as Development Infrastructure

Part IV · Shipping

Memory, Thermals, Battery

Privacy Architecture and App Review

Evals Without a Cloud

Plus three appendices: a measured model catalog, prompt patterns for 3B-class models, and the complete GenerationError troubleshooting table, enumerated by the compiler rather than copied from docs.

Pricing

One-time purchase. Quarterly updates included.

Book $59

12 chapters + appendices

PDF & EPUB, DRM-free

All future updates

Buy the book

Book + Code $99

Everything in Book

Companion repo: DigestKit, ModelBench, LocalEval

The full benchmark dataset

Buy book + code

Team $199

Everything in Book + Code

License for up to 10 engineers

Priority errata answers

Buy for your team

30-day refund, no questions. If the compile gates in Chapter 3 don't convince you, ask for your money back.

Free sample

Chapter 3: your first on-device feature, end to end.

Availability handling, typed guided generation, streaming into SwiftUI, and the three errors that dominate production. The complete lead chapter, free. It's the same feature the companion demo app ships.

Get Chapter 3 (PDF) EPUB

Author

Written from a home rack, not a content calendar.

Written by Roberto Gutierrez, founder of Digital Foundry and a staff iOS/tvOS engineer with more than twelve years on Apple platforms, who runs local inference daily: Ollama on a Mac Studio M2 Ultra serving a LAN, MLX on Apple silicon, evals against golden sets before anything ships. The same person restores mechanical Nikon film cameras and modded iPod Classics, which is to say: someone who thinks broken abstractions should be opened, measured, and fixed. This book treats language models the same way.

Questions

Fair questions, straight answers.

Isn't this all going to change at the next WWDC? Some of it. That's why updates are free and the code pins exact dependency versions. The iOS 27 material (the LanguageModel protocol, multimodal prompts, Core AI) is already tracked in clearly-marked sections and lands as a free revision when the SDK stabilizes. I could read Apple's docs for free. What's the $99 for? The docs describe the APIs. The book covers what they don't: which stack to pick and why, the memory budget that jetsams your app twenty turns in, the App Review notes that prevent a rejection cycle, the eval harness that catches Apple swapping the model under you, and the API drift the docs haven't caught up with, verified by compiler, with receipts. Does it cover [some giant cloud model]? No. This book is about the models that run on the machine in your user's hand, the 1–8B class, and it's plain about what they can't do. Chapter 1 includes the decision framework for when the cloud is still the right answer. What do I need to run the code? macOS 26+, Xcode 26+, Apple silicon. The eval and benchmark tools run on any M-series Mac; the demo app needs Apple Intelligence enabled. An Ollama-capable Mac (any M-series, more RAM is better) unlocks the Chapter 9–12 workflow.