Research

Only what we measured ourselves goes here.

  1. Running LFM3 on a phone to see if it can replace LangDict's AI features

    We ran Liquid AI's LFM3-VL-2B on an Android phone with 7GB of RAM to test whether it could replace translation, deck generation, and OCR in the language-learning app LangDict. In the end, we decided not to ship it.

  2. Lexa — a cross-card in-context spaced-repetition model

    Retention-prediction accuracy, against the default settings of FSRS-7

What research is for

At mibuki, research exists to solve the apps' problems and make the apps better. Rather than one large model for everything, we build models and methods narrowed to a single problem, measure them, and publish the results with their conditions.

  1. Find the problemProblems that show up while the apps are in use become research topics.
  2. Build and measureWe develop technology that solves the problem and test it against a fixed baseline under stated conditions.
  3. Ship it in an appTested technology goes into the app, so the people using it get it.

What it takes to be published here

We do not publish the conclusion alone. The baseline, the data and the metric sit in the same place as the number — a figure without its conditions can only ever flatter us.

A result is ready when a reader could follow the same steps and see the same thing. The measurement date and the version measured are always stated.

If we cannot produce a number, we do not claim a result. That is a constraint on us, not a feature.

The full standard (Approach, principle 03) →

Last updated: July 2026