The article analyzes the codebase of Kuzu, an embedded graph database acquired by Apple, highlighting architectural decisions like factorized execution and worst-case optimal joins that support Apple's on-device AI strategy.
# I Analyzed 163K Lines of Kuzu’s Codebase. Here’s Why Apple Wanted It.
Source: [https://medium.com/data-science-collective/i-analyzed-163k-lines-of-kuzus-codebase-here-s-why-apple-wanted-it-12294a7035fa](https://medium.com/data-science-collective/i-analyzed-163k-lines-of-kuzus-codebase-here-s-why-apple-wanted-it-12294a7035fa)
## *How a 10\-person startup built the graph engine Apple’s on\-device AI strategy was missing\.*
[](https://medium.com/@deshpandetanmay?source=post_page---byline--12294a7035fa---------------------------------------)
Press enter or click to view image in full size
Source — AppleI spent the last week reading through 163,000 lines of C\+\+20 that make up Kuzu, the embedded graph database Apple quietly acquired in October 2025\.
The acquisition surfaced through an EU Digital Markets Act filing — 10 employees, Waterloo Ontario, GitHub archived the same day\.
Every tech outlet ran the same angle: will Apple add a database to iWork? That reading misses what’s actually in this codebase\.
By the end of this piece, you’ll understand**three architectural decisions**— \* factorized execution, worst\-case optimal joins, and an Umbra\-style storage engine — that make this acquisition strategic rather than accidental\.
These aren’t incremental improvements over Neo4j or other graph databases\.
They’re fundamental design choices that solve the exact problem Apple Intelligence faces: reasoning about relationships across your entire digital life, on\-device, without a server\.
## What’s Actually in the Codebase
Kuzu is an embedded graph database\.
This article provides an in-depth review of ZCode, a desktop coding agent by Z.ai powered by the GLM-5.3 model, evaluating its features, workflow, cost, and performance on real projects.
This paper presents a reverse-engineered account of Apple Neural Engine (ANE) architecture, programming, and performance across A11-A18 and M1-M5 chips, based on direct measurements and static analysis.
A daily roundup of AI engineering news covering Codex's rapid user growth, Anthropic tokenizer cost issues, Apple's upcoming high-memory chip, and studies on tool adoption and KV-cache compression.
Apple's Lily engine optimizes on-device LLM inference for Apple silicon by leveraging unified memory and hardware, outperforming MLX-LM, and is tuned for the Qwen3.6-35B-A3B model's architecture.
The article retrospectively reverse-engineers Apple's Neural Engine on the M1 chip, analyzing its architecture and discussing its evolution with the M5 chip's integration into GPU cores.