Rituda builds two things: speech recognition for Azerbaijani, and a 27B language model rebuilt around it — a new tokenizer, continued pre-training, and the full post-training stack. Both run inside your own infrastructure.
Azerbaijani is under-served by the large providers — the accuracy is poor, the terminology is wrong, and the audio leaves the country. We build for this language specifically.
Azerbaijani speech to text
Speech recognition built for Azerbaijani rather than adapted from a multilingual model, and deployed the same way as the language model — inside your own network, on your own hardware.
27B dense, rebuilt from Qwen3.6-27B
Not a prompt wrapper and not an adapter on top of someone else's model. We replaced the tokenizer with one built for Azerbaijani and retrained the full parameter set through continued pre-training and the whole post-training stack.
Three stages, in this order. Nothing here is a wrapper around someone else's API.
The base model ships with a 248,320-token vocabulary learned mostly from other languages, and it cuts Azerbaijani into far more pieces than the language needs. We replaced it with a unigram tokenizer trained on Azerbaijani, which segments close to syllable boundaries and brings the text down to 2.3 tokens per word. Shorter sequences mean more text in the same context window and less compute for the same amount of language.
Swapping the tokenizer means the model has to relearn how to read. Continued pre-training on Azerbaijani, with supervised fine-tuning folded into the same stage, and preference optimisation last — teaching the model what counts as a good answer and what does not. Every stage runs across the full parameter set at 16-bit precision. No LoRA adapters bolted on the side, no quantised shortcut.
The base, Qwen3.6-27B, is published under Apache 2.0 — commercial use, modification and redistribution are all permitted. What we hand over is a derivative of it, delivered with its licence and a record of what changed. It runs on your servers and keeps running without us.
Both products are built to be deployed inside your perimeter. Nothing is sent to an external API, nothing is logged outside your network, and nothing depends on us staying online.
On-premise GPU servers, your own private cloud, or a fully air-gapped network — whichever your compliance position requires. Because the base model is Apache 2.0, the weights we hand over come with a licence that does not expire when the contract does.
Rituda has been building software in Baku since 2006 — platforms, applications and infrastructure for businesses across Azerbaijan. The AI work grew directly out of that: the same team, the same servers, pointed at the problem of making models actually work in Azerbaijani.
The people you meet are the people who train the models. There is no research division to hand you off to and no partner doing the real work behind the scenes — which is also why this page lists two products rather than seventeen.
Tell us what you have — hours of audio, a document archive, a support history — and what you want out of it. We will come back with an honest view of whether it is enough and what it takes to get to a result.