By Leo Marchandon and Gianluca Lo Nostro
July 7 (Reuters) - Ukraine will favour AI systems it can run on its own servers, a senior ministry official said on Tuesday, as wartime Kyiv seeks to keep digital tools for government services, businesses and the military from depending on remote systems that providers can restrict or switch off.
The approach favours self-hosted, or "on-premise," models that Ukraine can deploy on its own infrastructure, while limiting solutions that, by design, remain under the provider's control — a category that includes Anthropic and OpenAI's main models.
The policy was reinforced after the U.S. government ordered Anthropic to cut access to powerful models, echoing broader European sentiment, Roman Kyslyi, Chief AI Officer at Ukraine's Ministry of Digital Transformation, told Reuters.
"It confirms that AI sovereignty isn't just a defensive talking point, it's a necessity," he said.
Reuters reported on Tuesday that Chinese authorities are considering curbs on top AI models, which currently dominate the open-source market.
Kyslyi said the decisive criterion is not about where the model is from. "If the vendor will provide it to run on our on-premise (infrastructure), there are no restrictions."
"The model is essentially a commodity," Kyslyi said, adding that Ukraine would work with any provider whose technology could be deployed under Ukrainian control.
Currently, Ukraine's AI assistant inside the Diia government app runs on Google's remote-only Gemini, accessed through servers within the European Union. Kyslyi said Google provided free tokens for it, meaning no budget spending.
Still, Ukraine strips personal data before sending queries to Gemini because they "don't control those models," he said, describing Gemini as an "interim" solution.
Ukraine is also developing its own model with Kyivstar based on Google's Gemma, its open variant, due to be released in autumn and intended for use across government services, private enterprises and the military.
Kyslyi said the ministry compared several open-source options before choosing Gemma, including Mistral models and OpenAI's GPT-OSS. Gemma and Mistral matched remote-only alternatives on many performance tests, he said.
(Reporting by Leo Marchandon and Gianluca Lo Nostro in Gdansk, Editing by William Maclean)