Developers can enable context-aware Japanese input on macOS and Linux
Imagine typing kana and having it convert to kanji that fit the surrounding context without needing spaces. Karukan is a Japanese input system for Linux and macOS. It uses neural network models to handle complex context and remembers user selections to suggest better candidates for future input.
Neural network conversion and real-time feedback
Karukan performs conversion by inferring models based on GPT-2 or Qwen3 using llama.cpp. This goes beyond simple dictionary matching by analyzing surrounding text to enable natural kanji selection. Conversion results are displayed in real-time as you type, and conversion proceeds without pressing the spacebar. This feature can be toggled with the Ctrl+Shift+L key. User-selected conversion results are learned, increasing their priority in subsequent inputs, and learned candidates are presented during input.
Candidate generation and emoji input
It includes a candidate rewriter ported from Mozc, which automatically generates various notations for half-width katakana, English capitalization, and numbers (such as kanji numerals, Roman numerals, and hexadecimal).
- Candidate rewriter Automatically generates and annotates notations for half-width katakana, English capitalization, and numbers.
- Emoji input Emojis can be entered using kana reading methods and Slack-style trigger queries.
- Conversion persona Registering keywords in the configuration file enhances conversion for that specific topic.
Installation and initial model download
On Linux, it installs using the fcitx5 frontend, while on macOS, it is based on Swift and InputMethodKit. Specific installation methods for each platform should be referenced in the README of the respective subdirectory. On first run, required models are downloaded in the background from Hugging Face. Kana input and basic dictionary conversion are possible during the download, and neural network conversion activates automatically once model loading is complete. Subsequently, it uses the locally stored model without a network connection.
License and pre-use checks
Karukan is provided under a dual MIT or Apache-2.0 license, while some data is derived from Mozc and follows the BSD 3-Clause license. Because an initial model download is required for neural network conversion, a network connection is essential on first run. Conversion accuracy for specific industry terms or proper nouns may vary depending on training data and persona settings, so it is advisable to review the candidate list thoroughly before writing important documents. Additionally, compatibility in fcitx5 environments may differ across Linux distributions, so stability in the specific environment should be verified before use.
