InfinilangInfinilang
Accelerated Japanese, tailored to you.
Every learner hits the same wall: WaniKani teaches you five hundred kanji, and then there is nothing to read that uses those five hundred and nothing else. Infinilang generates it, forever.
macOS · Windows · Linux
Reading
Comprehensible input, computed rather than curated.
Infinilang pulls your live WaniKani assignment state and builds passages from your known-kanji pool, weighted toward the ones your SRS says are shakiest. Exactly one upcoming kanji is introduced per session and badged as new.
Furigana that get out of the way
Set a threshold and readings appear only above kanji you have not passed yet. Anything that slipped into the text from outside your level always keeps its reading, so you are never stuck.
Know what you just read
A coverage panel accounts for every character: how many of your known kanji the passage used, how many came from the next two levels, and anything outside your level that was auto-glossed. Hover any word for its meaning and readings.
Particles you have not absorbed
A tracker tokenises everything you read, watches which grammatical particles you demonstrably understand, and biases future generation away from the ones you do not.
Speaking
Every word you say, graded as you say it.
Free-form conversation where the transcript colours itself word by word from green through amber to red, so you can see exactly which syllable let you down.
Yuki
A casual friend in her twenties. Warm, and she teases.
Tanaka-sensei
A formal teacher. Reformulates what you said rather than correcting you.
Akari
A classmate with her guard up. 「べ、別に〜」
Sato-san
The neighbourhood shopkeeper. Polite, brisk, endlessly patient.
Seven places to be
A cafe in Shibuya, a Tokyo train station, an office job interview, a doctor's visit, a department store, lost in the city, and a language exchange. Each carries an expected register, so casual, polite and keigo get practised where they belong.
It teaches without telling you off
A corrector spots your mistakes and hands them to a planner, which works them into what your partner says next. You get taught to rather than called out, and a sidebar quietly offers more natural ways to have said it.
On the screen
This app has not shipped yet. Captures land here at release.
Memory
It remembers you, and you can edit it.
After every turn a fast model extracts one durable fact about you, rates how important it is, and stores it locally with an expiry. At the end of a session a reflection pass consolidates those into higher-level insights. Next time, the ones that matter are retrieved and seeded into the conversation, so your partner picks up where you left off.
All of it is visible in a drawer, and every fact can be read, edited or deleted. A language partner that remembers your job and your dog is a good thing only if you can see what it thinks it knows.
Your keys, your machine
Nothing here is a service.
Infinilang has no account, no server and no billing. Point it at a model endpoint you run, at Ollama on the same laptop, or at Anthropic with your own key. Embeddings and the vector store are local: memory vectors sit on the fact's own SQLite row and are searched in process, so there is no vector database to sign up for.
Locked down by construction
Keys live in the operating system keyring. A strict content security policy pins an explicit allowlist of hosts the app may contact, so nothing else is reachable from it even in principle. There is no analytics and no telemetry.
Made to be readable
Motion, contrast and text size are all adjustable, there is an OpenDyslexic option, and the whole interface honours a reduced motion preference. The app itself is available in English, Spanish, Chinese and Korean.
Detail
The specifics.
Infinilang is early. Reading mode is the mature half; speaking mode is a working preview and is still being built out. It is listed here because it is real and usable, not because it is finished.
- Platform
- macOS 11 or later, Windows, Linux. Mobile is a later phase.
- Status
- Early development, version 0.1.0. No packaged release yet.
- Pricing
- Free. Bring your own API keys, or self-host the models.
- Licence
- Apache-2.0
- Teaches
- Japanese. Interface in English, Spanish, Chinese and Korean.
- Models
- Any OpenAI-compatible endpoint, Anthropic, or Ollama locally
- Speech
- Azure Speech for recognition and synthesis, with per-word pronunciation scoring
- Privacy
- No accounts, no telemetry. Local SQLite database and vector store; keys held in the OS keyring.
Questions worth asking.
- What does Infinilang cost?
- Nothing. It is Apache-2.0 licensed and there is no billing code in it. You supply your own API keys, or point it at a machine you already run, and you pay whoever provides those services rather than paying us.
- Do I need a WaniKani subscription?
- For the reading half, yes: the whole idea is that passages are constrained to the kanji your WaniKani account says you know, weighted by how well you know them. The speaking half works without it.
- Can it run entirely on my own hardware?
- Almost. The default provider is a self-hosted OpenAI-compatible endpoint, Ollama is a first-class option, and the vector store is plain SQLite on your disk rather than a hosted database. Speech recognition and synthesis currently go to Azure, which is the one part that must leave the machine.
- Which languages does it teach?
- Japanese only. The interface itself is available in English, Spanish, Chinese and Korean.