Veon-owned Kyivstar and Ukraine’s Ministry of Digital Transformation jointly selected Google’s Gemma foundation framework to develop a sovereign national large language model (LLM).

Under the initiative, Kyivstar and Ukraine’s WINWIN AI Center of Excellence will develop a sovereign Ukrainian LLM designed to reflect the country’s language, dialects, history and cultural contexts while ensuring sensitive data is stored and processed within the country.

Google’s open-source LLM was selected following what the partners described as an extensive technical evaluation with an emphasis on national security. Danylo Tsvok, chief AI officer at the ministry and WINWIN CEO, said the selection process focused on “how well it already handles Ukrainian-language texts and how controllable it is during additional training” to “minimise linguistic and ethical risks in our LLM”.

“Choosing Gemma for the Ukrainian LLM gives us the best balance between performance and resource use, as well as high-quality training,” explained Mykhailo Nestor, Kyivstar’s director of digital product development. He highlighted Gemma’s support for more than 140 languages including Ukrainian, its ability to handle large volumes of information in a single interaction and multimodal support for both text and image-based tasks. Flexibility in model sizes was also highlighted as an advantage for tailoring deployments to different sectors.

Gemma has already been used as a base for Ukrainian-language models MamayLM and Lapa LLM from open source AI community Hugging Face.

Kyivstar will now work to optimise Gemma for Ukrainian language and train the model on curated national datasets, supported by dedicated benchmarks. Once live, the LLM is expected to support AI-powered services across sectors including government, education, healthcare and finance.

The partners added that Google’s selection further strengthens the country’s economic ties with the US following Kyivstar’s listing on the Nasdaq stock exchange earlier this year.