An open-source neural IME for macOS, alpha honesty included

What it does

azooKey-Desktop is a Japanese input method editor for macOS that converts kana into kanji using a local neural engine called Zenzai. Written in Swift, it behaves as a system input source and offers live conversion, history learning, and user dictionary support. The authors are upfront that the project is alpha software and offer no operational guarantees.

The interesting bit

Most IMEs are opaque infrastructure you tolerate; this one ships with an optional LLM-driven “Magic Conversion” feature and plugs into a separate personalization backend called Tuner to adapt to individual writing habits. The README also nods to community forks for Linux and Windows, suggesting the engine is portable enough to outgrow its macOS roots.

Key highlights

• Local neural kana-kanji conversion via the Zenzai engine, with profile prompts and history learning
• Optional LLM-based “Magic Conversion” for rephrasing suggestions
• Integration with the Tuner system for per-user personalization
• Native AZIK romanization support
• Explicitly alpha: macOS 15 is confirmed, while macOS 14 and 26 remain unverified

Caveats

• The project is self-described alpha software with no guarantees of stable operation
• macOS 14 and macOS 26 support is listed but explicitly unverified
• Neural model weights are distributed via Git LFS, and the README warns that missing or pointer-only weight files silently degrade conversion quality

Verdict

A solid curiosity for Japanese-speaking macOS users who want a hackable, locally running alternative to Apple’s IME. Avoid if you need battle-tested stability or prefer your system input methods to stay out of Git LFS territory.