Skip to content
Czech LLM family

From scratch. In Czech.

Three language models on APPMAXX's own architecture - Tiny, Small, and the Mid flagship. Every weight trained from zero, proudly Czech to the last one.

3
sizes in one Czech family
67–974M
params - pocket to flagship
Custom
home-grown architecture
🇨🇿
Czech-first, built in Prague
Playground

Small talk. Literally.

Ask it for a haiku, a questionable joke, or the one true best beer. Miracles not included. Charm very much included.

Open the playground
Models

Three sizes.
One family.

Flagship Almost here

MAXX Mid

974M parameters

The big one. Nearly a billion parameters on a trillion-token Czech run - our flagship, built for the heavy lifting the little ones leave on the table. Same home-grown architecture, scaled all the way up.

974M
parameters
1T
token run
24 × 1792
depth × width
In the making - follow ai.appmaxx.io
In training

MAXX Tiny

67M parameters

Our featherweight - a 200-billion-token Czech run, in training right now. It fits on a laptop, on a Raspberry Pi, and very nearly on a pocket calculator.

Almost here

MAXX Small

217M parameters

The middle child - roughly 3× Tiny's muscle on a 500-billion-token run. Still small by the world's standards. Just right by ours.

Next in line - follow ai.appmaxx.io
Under the hood

From scratch. Literally.

No fine-tuning. No distilling someone else's model. Every MAXX model stands on APPMAXX's own architecture, and every single weight is born right here - from the tokenizer to the last training step, on a Czech-first, quality-over-quantity data mix.

Why? Because we wanted to understand every last bit ourselves. And because small models are the future everywhere the giant ones are overkill.

Architecture APPMAXX custom (from scratch)
Parameters 67M · Tiny / 217M · Small / 974M · Mid
Training run 200B / 500B / 1T tokens
Language Czech-first
Origin Prague, CZ 🇨🇿
The culprits

A software studio in Prague that went and built its own LLM. Because why not.

By day we ship SaaS, web apps and MVPs. MAXX is our detour into language models - and we're only just warming up.