Why we train our own models

Here's the uncomfortable part of most “personalized learning” apps: they're a ChatGPT call in a trench coat. Wrap an API, bolt on a streak counter, call it personalized. Lernex is the other thing — we're actually training the models that do the teaching.

Off-the-shelf models don't fit the job. The big proprietary ones are too slow and too expensive to run on every lesson, hint, and quiz a learner touches. The open ones are cheap but built to be everything to everyone, so they're rarely great at the one thing we need. So we build our own.

Metis is that model line: trained from scratch on Lernex's own data and schemas, designed to do one job absurdly well — teach a single person. Not the average student. You.

We're after hyper-compact, efficient models and the exotic architectures the frontier labs are too cautious to ship — small enough to run constantly, sharp enough to genuinely tutor, weird enough to get us somewhere the wrappers can't follow. And we open-source all of it. Try it, break it, tell us where it's wrong. That feedback is the point.