In Kazakhstan, the first national language model based on artificial intelligence, IrbisGPT, has been published in the open access.

In Kazakhstan, the first national open-source language model, IrbisGPT, trained on a large dataset in the state language, has become available. The developers published the official release on the popular platform Habr.com. Now, every Kazakhstani can test the language model and contribute to its training in the Kazakh language.

IrbisGPT is a public non-profit initiative developed in collaboration with MOST Holding and Gen2b.ai, a studio specializing in applying artificial intelligence to business.

Compared to the pre-release version published in the spring, the current version of IrbisGPT demonstrates excellent potential for training. According to the developers, similar open-source models either attempt to answer in English or simply "spout" random words in Kazakh, but thanks to extensive training, the national language model provides sufficiently detailed and correct answers to questions without context. For example, IrbisGPT answers questions in the state language such as "why is grass green," knows who the President of Kazakhstan is and how many days are in a year, and can even philosophize about the meaning of life.

The final tokenizer vocabulary contains more than 60,000 tokens. The team has a plan to create a model with a more advanced architecture that will be useful in various industries.

You can download the pre-trained model via the link.