KanaAI
現在の製品・配布先は Kotori日本語入力 です。
紹介サイトはKotoriの現行情報を案内します。
以下のKanaAI実装・beta.2検証記録は旧プロジェクトの履歴です。
Mozcを基盤に、ローカルAIを育てるWindows日本語IME。
A local-first Windows Japanese IME built on Mozc, Rust, and bounded local AI.
Windows TSF · Mozc baseline · Local AI · No cloud dependency
[!IMPORTANT]
Status: public beta published - unsigned, with a bundled local AI that now
actually starts.
Download: v0.1.0-beta.2
(KanaAI-0.1.0-Setup.exe / KanaAI-0.1.0-x64.msi, Windows x64, about
1.1 GB - the model is in the package).
Japanese introduction and installation guide: KanaAI website.
This repository contains the public source and engineering record.
The Release body records the evidence and limitations of each published binary.
The product is a native Windows TSF text service, not the retired
Workbench/CLI demo. Do not treat a local build, a DLL load, a source test, or
a loopback model mock as proof that the end-user IME is ready.
What beta.2 established, and what it did not. The installer lifecycle is
machine-verified by receipt for the published hash (11/11 phases), Japanese
input is operator-confirmed with the text service observed loaded in eleven
real processes, and the bundled AI is measured starting from the package's
own opt-in on the product path - a 1.1 GB model loaded into a child process
at a 1.6 GB working set, with byte-hash verification and a completed
inference. What is not established is that any of that improves a
conversion: the published beta.2 AI-on/AI-off measurement found no change
in conversion results, and the last
recorded model-quality evaluation in STATE.md was not shippable
in any of its six roles. Read the Release body before installing.
Everything still runs offline. The runtime listens on loopback only, behind a
per-process key, and performs no network access.
This beta is not a completed product and .goal-complete has not been
created.
Windows release contract ·
Product requirements ·
Current engineering state ·
Open work and release checklist ·
Local AI runtime plan ·
Security ·
Contributing
What KanaAI is
KanaAI is being built as a practical Japanese input method for Windows:
- Mozc owns everyday input. Romaji composition, segmentation, conversion,
dictionary behavior, candidates, preedit, and commit remain owned by the
pinned Mozc engine.
- Rust owns bounded orchestration. A private broker tracks sessions and
generations, validates results, and keeps optional AI work away from the
synchronous key path.
- Local AI improves candidates when it is safe. The intended first bundle
uses a pinned Qwen GGUF and a CPU-only
llama.cpp runtime. A model result can
only reorder a bounded, unchanged candidate set; it cannot invent text or
take ownership of commit.
- Mozc remains the fallback. Missing model, timeout, malformed output,
process failure, protected field, or low resources must leave normal input
usable.
The current implementation target is a Windows x64 native TSF TIP. A
browser Workbench, HTTP API, CLI, or source bridge is development tooling—not
the product.
Product principles
| Principle | What it means |
|---|
| Native first | The release target is a registered Windows TSF text service used in ordinary desktop applications. |
| Fast path stays fast | Key and preedit callbacks do not wait for a model, disk I/O, or network I/O. |
| AI is bounded | At most a small candidate window is reranked; IDs, text, and commit authority are validated. |
| Local by default | The intended release runs its model/runtime locally and does not require a cloud account or remote endpoint. |
| Explainable fallback | A missing or rejected AI result returns the Mozc baseline instead of silently rewriting text. |
| Privacy is a release gate | Protected/password contexts, bounded context, no user-data logging, and no hidden telemetry are required. |
Current status
This table describes evidence, not marketing claims.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
