Ukraine is advancing its technological capabilities by developing a large language model based on Google’s open-weight Gemma framework, with the goal of establishing a fully independent national AI system for both military and civilian applications. The initiative, led by the digital ministry in collaboration with mobile operator Kyivstar, will initially utilize Google’s computing infrastructure for training before moving to Ukrainian data centers, ensuring the country maintains complete control over the system that will serve 23 million citizens.
Officials indicate that the new AI model is intended to enhance a range of services across public administration, business, and defense sectors. Oleksandr Bornyakov, Ukraine’s Deputy Minister for Digitalisation, noted that the country aims to avoid spending millions on foreign models and to decrease reliance on proprietary systems such as OpenAI’s ChatGPT. Reports from News.Az, citing Reuters, reveal that alternatives like Chinese models DeepSeek and Qwen were also dismissed during the selection process.
Ukraine has already integrated various AI tools into its military operations, including systems provided by U.S. firm Palantir for tasks like reconnaissance and strike analysis. The decision to adopt Google’s Gemma was influenced by its multilingual capabilities and favorable policy features, according to Kyivstar. Other contenders, such as Meta’s Llama and France’s Mistral, were reviewed but ultimately not selected.
This initiative specifically addresses communication challenges that existing AI systems often face. Many Ukrainians communicate using blended dialects incorporating Ukrainian, Russian, and minority languages. The chief product officer at Kyivstar identified ongoing issues such as mistranslations and legal inaccuracies that the new model aims to rectify, as well as the phenomenon of AI “hallucinations” that can undermine trust in automated systems.
To ensure the AI effectively represents Ukrainian and minority languages, including Crimean Tatar, four advisory committees with binding oversight have been established. These committees will guide the AI’s technical, legal, cultural, and linguistic development. Data is currently being collected from over 90 institutions, encompassing court records, educational materials, and archives related to the ongoing conflict.
Initially, the model will be trained on secure GPUs located abroad due to the risk of Russian military strikes. Once developed, it will be deployed within Ukraine, where officials expect significant cyber threats upon the system’s launch. They are proactively preparing defenses against potential risks, including prompt-injection attacks. In anticipation of energy disruptions, Kyivstar has already installed thousands of backup generators to maintain operational stability during intensified strikes.
Once completed, the AI model will first support government services and Kyivstar platforms, with plans to expand its capabilities to the broader private sector. Ukrainian officials assert that this project exemplifies how smaller nations can leverage open technologies to achieve strategic autonomy in the field of artificial intelligence.
