In Ukraine, a large language model called Lapa LLM was presented, which was the result of collaboration between scientists from several leading institutions: the Ukrainian Catholic University (UCU), the Kyiv Polytechnic Institute (KPI), the Lviv Polytechnic , and the Academy of Mining and Metallurgy in Krakow . According to the press release, the model was based on Google 's Gemma-3-12B .
The model name — Lapa LLM — was chosen in honor of Valentyn Lapa , who previously co-authored methods of group argument accounting — one of the theoretical foundations of modern deep learning. According to the developers, during the creation of the model, about 80,000 tokens were replaced from the initial 250,000 to better adapt the architecture to the Ukrainian language.
UCU representatives report that in terms of speed and quality of work with the Ukrainian language, Lapa LLM surpasses the original Gemma, as well as most closed models of the same class. This advantage, they claim, was achieved through localization, tokenizer optimization, and adaptation to Ukrainian morphology and syntax.
This news is important not only for the scientific community, but also for the development of the Ukrainian IT market, since models of this class are a key tool for natural language processing, translation, text creation, chatbots, automated data analysis, and other applications.
In the near future, open testing of Lapa LLM is expected, as well as the integration of the model into projects related to the Ukrainian language, culture, education, and business.
