Emvryx’s first models, Ember 2 and Flare 2, were trained from scratch: their own tokenizer, transformer and training run. The 2.x series builds on open Qwen models (Apache 2.0) and is fine-tuned here on Emvryx’s own lessons: looking things up and showing the source, saying “I don’t know”, following your rules. None of ChatGPT’s, Claude’s or Gemini’s answers are in its lessons. Some practice lessons are written by an open model and checked by rules.
We wrote the tokenizer, the transformer and the training loop, then trained it on carefully chosen reading: educational pages, everyday web text, all of English Wikipedia, code in many languages and step-by-step maths.
It answers from its notes and names the source. It says “I don’t know” instead of guessing, rechecks when you say it’s wrong, keeps to exact counts and formats, follows a change of topic, and reads through typos.
Every new version keeps everything the last one learned, and must beat it on the same tests before it replaces it. Bigger siblings follow, and the app stays the same.
Any Emvryx model can run wherever you want it, and be shared as widely as you want. Same models, same app; only the address and the guest list change.
Like any AI, Emvryx can make mistakes. That is why it shows its sources and which model answered, so you can check what matters, especially for homework, health and money.