Kyivstar, Google to build sovereign Ukrainian LLM

Ukraine is building its own national large language model, using Google’s open-source Gemma system and Google Cloud’s Vertex AI as the core technology. Mobile operator Kyivstar will lead the project with the Ministry of Digital Transformation and its WINWIN AI Centre, training the model on curated Ukrainian data so it captures local dialects, terminology, and history. The partners say all sensitive data will be stored and processed inside Ukraine, with future use in government, healthcare, finance, and other regulated sectors.

Engineers will adapt Gemma specifically for Ukrainian, including tokenizer work and new benchmarks for fine-tuning. Officials say they chose Gemma after extensive testing, based on how well it already handles Ukrainian and how controllable it is during further training, aiming to reduce linguistic and ethical risks. The project also seeks to counter Russian influence in existing Ukrainian models, which often rely on data shaped by Russian narratives. VEON, Kyivstar’s parent company, describes the effort as part of a broader push for local-language AI across its markets, including earlier models in Kazakhstan and Pakistan.