Ukraine’s Approach to Sovereign AI in Public Administration

Why governments are rethinking language models

As artificial intelligence becomes embedded in public administration, governments are increasingly confronting a foundational question: should core language technologies be treated as external tools, or as public digital infrastructure? Large language models are no longer confined to innovation labs. They shape how laws are interpreted, how services are delivered, and how citizens interact with the state.

For Ukraine, this question has immediate operational consequences. Reliance on foreign AI systems has exposed limitations in language support, contextual accuracy, and institutional alignment. Most global models struggle with Ukrainian legal terminology, administrative processes, and historical context. More importantly, externally developed models are trained on global datasets that may contain factual distortions or political bias, particularly in contested geopolitical environments. For public institutions, this creates risks that go beyond technical performance and affect trust, accountability, and policy coherence.

Ukraine’s rationale for a national LLM

Ukraine’s decision to develop a national Large Language Model (LLM) is rooted in capacity, not prestige. The initiative is led by the Ministry of Digital Transformation as part of a broader strategy to strengthen the state’s ability to deliver services digitally while maintaining control over critical data and technologies.

Recent independent research has shown that many open-source language models reproduce misleading narratives about Ukraine when responding to factual queries. These outcomes are rarely intentional, but they demonstrate how training data shapes automated interpretation. As language models are increasingly embedded in search, translation, and document analysis tools, such distortions can scale quickly across public and private systems. For a government, this makes full reliance on external models a structural vulnerability rather than a neutral technical choice.

The national LLM is intended to mitigate this risk by anchoring its own language model in verified Ukrainian sources, legal frameworks, and institutional realities. Its primary application will focus on powering Diia, the educational app Mriia, and critical defense services.

“Ukraine is on a mission to become one of the top three global leaders in AI development and implementation by 2030. We are already transitioning toward an 'Agentic State,' where AI-driven assistants proactively meet the needs of our citizens and government. However, true progress in government and defense is impossible without AI sovereignty. We build our national LLM, intended to run on our sovereign AI Factory infrastructure, to ensure that our critical data stays within the country and that the AI products for citizens remain free from propaganda and disinformation”, said Danylo Tsvok, Chief AI Officer at the Ministry of Digital Transformation of Ukraine and CEO WINWIN AI Center of Excellence.

Institutional governance and the role of the AI Center of Excellence

The development of the national LLM is spearheaded by the WINWIN AI Center of Excellence. It is the key government body responsible for AI development, launched by the Ministry of Digital Transformation in February 2025. The Center serves as the initiative's institutional owner and is responsible for aligning technical development with public policy objectives.

To ensure the depth of the national LLM, the AI Center of Excellence has decentralized its development across specialized domains. Building a sovereign AI is far more than a technical challenge; therefore, the project integrates high-level expertise in ethics, linguistics, culture, and history. The government engaged professionals from academia and the business community. A working group has been established to oversee the model’s quality, namely, by creating rigorous performance benchmarks.

The technical foundation focuses on model architecture, large-scale deployment, the maintenance of robust computing infrastructure and secure datasets. From an ethical standpoint, the project embeds a framework that guarantees respect for human rights, data protection, and adherence to European legal standards. To safeguard national identity, cultural and historical experts meticulously filter inputs to eliminate disinformation while ensuring the model reflects a true social context. Finally, the linguistic direction ensures the LLM masters the nuances of the Ukrainian language, from technical terminology and phraseology to local dialects.

Public–private delivery under state control

Implementation is carried out in partnership with Kyivstar, Ukraine’s leading digital operator. Rather than outsourcing the initiative to a global AI provider, the government opted for a domestic infrastructure partner with experience operating large-scale systems under high-risk conditions.

“Participation in the development of a national Ukrainian LLM is a natural step for Kyivstar as a technology leader, since our strategic objectives fully align with the government’s vision for advancing the AI sector. Together with the Ministry of Digital Transformation of Ukraine, we aim to build an open national ecosystem where business, government, and academia will jointly develop Ukraine’s AI infrastructure to ensure the country’s technological sovereignty and support national digital services,” said Oleksandr Komarov, CEO of Kyivstar.

Kyivstar finances the initial development phase and is responsible for assembling engineering teams, managing cloud-based computing resources, and supporting early-stage training and testing. The state retains ownership of the model, defines its purpose, and controls its future use. This arrangement reflects a pragmatic response to limited public resources and the current shortage of domestic AI infrastructure, while preserving public governance over the technology.

For governments considering similar initiatives, this approach illustrates how private-sector operational capacity can be leveraged without relinquishing strategic control.

Technical approach and infrastructure constraints

Ukraine’s LLM is being developed using a pre-training approach based on the open-source language model Gemma by Google. Training a foundation model from scratch would require computing resources that are currently unavailable domestically and would significantly extend timelines. Pre-training allows the team to focus resources on contextual adaptation rather than baseline language acquisition.

The target model size is intentionally moderate, prioritizing deployability and efficiency over scale. Properly trained mid-sized models can perform well on specialized tasks such as legal text analysis, document classification, and administrative support, which are central to public-sector use cases. The choice of architecture is guided by technical and policy criteria, including security considerations and licensing conditions.

“Ukraine’s experience shows the world that bold innovation is a prerequisite for resilience. When the public sector adopts the agility of the technology industry, a unique model of public-private partnership emerges – one where security is not a constraint, but the foundation for growth. At Google, we see Ukraine as a strategic partner whose approaches to AI governance and digital transformation are setting a new international benchmark,” noted Anna Bulakh, Government Affairs and Public Policy Lead for Ukraine, Google.

Data as a public-sector asset

The most resource-intensive component of the project is data preparation. Language models intended for public administration require high-quality, domain-specific corpora that reflect how language is used in law, policy, and institutional practice.

The Ministry of Digital Transformation is coordinating the creation of a large Ukrainian-language corpus drawn from legislation, court decisions, regulatory documents, academic research, historical archives, and cultural sources. Many of these materials exist across public institutions but have never been consolidated or structured for machine learning. Their aggregation requires inter-institutional coordination and clear legal and ethical frameworks.

Personal and sensitive data are explicitly excluded from training. All materials undergo anonymization and filtering processes, overseen by ethical governance bodies. The Ministry is also exploring mechanisms for voluntary contributions from authors and publishers, addressing broader questions around copyright and public interest in AI development.

Initial applications in government and public services

The national LLM is being developed with specific public-sector applications in mind. One of the use cases is the analysis and translation of European Union legislation, a task central to Ukraine’s accession process and currently dependent on limited human capacity. Automating parts of this workflow can significantly reduce administrative burden while improving consistency.

The national LLM will serve as the foundation for a launched Diia.AI, a national agentic AI assistant integrated directly into the Diia app and portal, Ukraine’s digital government platform. Language-based interaction could simplify access to services, support guided processes, and improve communication between citizens and the state.

“The concept of an ‘Agentic State,’ which we are jointly implementing through Diia.AI based on Google’s AI models, reflects a fundamental shift in the logic of public services. Governments are moving away from passive interfaces toward intelligent systems capable of autonomously executing tasks and anticipating citizens’ needs. Our goal is to help governments to enhance the capacity of every civil servant through AI tools, transforming the public sector into a flexible, human-centric system” said Anna Bulakh, Government Affairs and Public Policy Lead for Ukraine, Google.

Openness, reuse, and long-term impact

Following a beta testing phase involving public institutions and researchers, the model is expected to be released in open access. This decision reflects transparency, auditability, and ecosystem development. Open source allows other bodies, including businesses, to deploy the model locally, adapt it to specific needs, and assess its behavior independently.

In the longer term, the government anticipates that the national LLM will enable domestic businesses to develop Ukrainian-language AI tools aligned with public-sector standards. For the state, the primary measure of success will not be global competitiveness, but reduced dependency on external providers and improved institutional capacity to work with language and data at scale.

Ukraine’s approach demonstrates that sovereign AI in government does not require full technological self-sufficiency. It requires clarity about which digital capabilities are strategically sensitive, governance structures that reflect public accountability, and investment in data as a public asset. These lessons are transferable well beyond Ukraine’s context and increasingly relevant to governments navigating the integration of AI into core state functions.