A Japanese IME that runs a local GPT-2 model

What it does Karukan is a Japanese input method for Linux and macOS built in Rust. It converts romaji to hiragana and then to kanji using a GPT-2-based model executed locally via llama.cpp. The project bundles frontends for fcitx5 on Linux and InputMethodKit on macOS, backed by a shared engine that tracks state, handles live rewriting, and memorizes your preferred conversions.

The interesting bit Rather than leaning entirely on classical morphological analysis, it throws a local large language model at the problem of choosing the right kanji sequence. It also borrows Mozc’s candidate rewriter to generate variants like hex numbers and half-width katakana, and supports Slack-style :emoji triggers alongside phonetic kana input.

Key highlights

• Neural kana-kanji conversion via a local GPT-2 model running through llama.cpp
• Live conversion that updates candidates as you type without hitting space
• Context-aware suggestions that factor in surrounding text
• Learns from your selections and promotes them in future predictive lookup
• Candidate rewriter (from Mozc) auto-generates numeral variants, half-width katakana, and case changes; emoji lookup by kana or :trigger syntax

Caveats

• First launch downloads the neural model from Hugging Face, so expect a slow initial startup; subsequent runs use the cached model.

Verdict Developers and writers who type Japanese on Linux or macOS and want an offline, learning-capable alternative to cloud-based or traditional rule-based IMEs should take a look. If your current Mozc or system setup already feels invisible, the first-launch download and local model overhead may not be worth the switch.