# The Art of Core AI

> Field techniques for putting AI on your phone — a free book written entirely from real
> measurements: 30+ models (LLMs, MoE, diffusion LLMs, speech, music, vision) shipped to real
> iPhones and Macs on Apple's Core AI framework (iOS/macOS 27). Core thesis: on-device AI is
> governed by how many bytes you move, not how fast the chip computes.

## Book

- [Full book, single page](https://john-rocky.github.io/the-art-of-core-ai/): all 13 chapters + labs
- [Chapter sources on GitHub](https://github.com/john-rocky/the-art-of-core-ai): markdown per chapter
- [Japanese edition on Zenn](https://zenn.dev/mlboydaisuke/books/coreai-textbook): the original

## Topics covered

- Why memory bandwidth, not compute, decides on-device LLM speed (ch. 1, 2, 11)
- The Apple Neural Engine vs GPU trade-off, measured (ch. 3, 10)
- Quantization: what 4-bit actually breaks, int8/int4/QAT (ch. 4)
- Mixture-of-Experts on phones (ch. 5)
- KV caches, prefix caching, speculative decoding (ch. 6, 7)
- Prefill vs decode, chunked prefill (ch. 8)
- Diffusion LLMs and parallel decoding (ch. 9)
- Shipping: bundles, distribution, being a first-class citizen on the platform (ch. 12, 13)

## Author

- Daisuke Majima (MLBoy) — [github.com/john-rocky](https://github.com/john-rocky), [huggingface.co/mlboydaisuke](https://huggingface.co/mlboydaisuke), [x.com/JackdeS11](https://x.com/JackdeS11). The same person ports the model zoo, runs the devicemark leaderboard, and wrote this book.

## Related projects by the same author

- [Core AI knowledge base](https://john-rocky.github.io/coreai-model-zoo/): the measurements this book was written from, as a page per topic — stateful KV cache, AOT and specialization, ANE vs GPU authoring rules, custom Metal kernels. the measured behaviour behind each chapter, which Apple's API documentation does not cover
- [coreai-model-zoo](https://github.com/john-rocky/coreai-model-zoo): 81 models converted to Core AI, each gated against its source model, with the recipe that produced it
- [coreai-kit](https://github.com/john-rocky/coreai-kit): run any zoo model in one line of Swift; FoundationModels-compatible
- [awesome-core-ai](https://github.com/john-rocky/awesome-core-ai): curated Core AI ecosystem resources
- [Hugging Face models](https://huggingface.co/mlboydaisuke): downloadable .aimodel bundles