karukan

Why we included this project

Conventional Japanese input methods on Linux and macOS rely on dictionary-based conversion, but Karukan runs GPT-2 and Qwen3 models locally through llama.cpp and converts kana to kanji as you type, without waiting for you to press Space. It learns from your choices too: pick a candidate once and it comes up first next time, and it handles emoji and alternate number notations. The project is split into a shared core engine, frontends for fcitx5 on Linux and InputMethodKit on macOS, and CLI tools that build system dictionaries from SudachiDict data. That layout makes it a useful reference for developers who want to see how a neural IME is wired end to end, and for Japanese-speaking users it's a genuine alternative to proprietary input methods.