Know-how, not black boxes
Having trained a model and run a crawler ourselves, we can judge what an AI or search product really needs – and what it does not.
Lab
Technology we build for ourselves.
Besides games and client work, NETBYTE develops its own foundations: a language model for Europe's languages and an independent search engine. Neither is finished yet. This page shows where they stand.
In training
Language model
Helix
A language model with 320 million parameters, trained from scratch for all 24 official EU languages – on rented gaming GPUs instead of a data centre. We document every step, including the mistakes.
In development
Search engine
Mira
An independent web search with its own crawler and its own index. Mira respects robots.txt and sitemaps, waits politely between requests, filters out duplicates and ranks results by language and country. It is still an early prototype.
02
Why we do it
Three reasons to build it ourselves.
Europe's languages
Smaller languages get little attention from the big models. Helix treats all 24 official EU languages as first-class from the start.
Measured in public
We publish numbers, costs and mistakes on the blog. That keeps us honest, and others can learn from it.
Shared interest?
Building something with AI or search?
Whether you need a search for your own content or want to know what a small language model can do: tell us what you are planning.