Scientific Intelligence How Russian AI Models Are Learning to Understand the Language of Science
Researchers at HSE University have developed an approach for adapting large language models to work with Russian-language scientific terminology.
Scientists at HSE University have taken an important step in the development of domestic artificial intelligence. They developed an approach for adapting large language models to work with Russian-language scientific terminology. The results are notable: the adapted model operates 2.7 times faster and requires 73% less memory than the original open multilingual system. The software has already completed state registration.
The key objective behind the project is to train AI systems to analyze Russian scientific texts, patents, reports and analytical materials more accurately. That requires more than grammar and syntax alone. The system must also understand professional context, including scientific and technological terminology, relationships between concepts and industry-specific formulations. The foundation of the model is the iFORA-QA dataset, manually assembled by more than 150 experts from HSE ISSEK’s Institute for Statistical Studies and Economics of Knowledge using materials related to science, technology and innovation.
Efficiency as an Advantage: Fewer Resources, More Possibilities
At a time when computing power remains an expensive resource, lowering hardware requirements is not merely a technical achievement, but a strategic advantage. The HSE model can run on more accessible hardware, potentially reducing the cost of AI deployment for universities, research centers and organizations that lack large-scale GPU infrastructure.
That is particularly significant for Russia because the project aligns with one of the country’s central AI trends: the creation of specialized domestic language models and applied AI agents for closed professional environments. Unlike general-purpose systems, this model is tailored to Russian-language sources, Russian scientific terminology and the country’s science and technology agenda.
What Could Science and Society Gain?
The development could accelerate both scientific and applied research, speed up the processing of patents and scientific publications and improve the quality of expert analytics. Over time, that may help shorten the path toward new medical, engineering, educational and industrial solutions.
For Russia, the project strengthens technological sovereignty in AI and scientific analytics. This is not another general-purpose chatbot, but a specialized tool for scientific and technical analysis that could be deployed across universities, research institutes, government agencies, corporations and R&D divisions.
Smart Search, Knowledge Graphs and Multi-Agent Systems
The next stage involves turning the adapted model into a comprehensive scientific and technological analytics system. As early as 2026, researchers plan to build additional tools on top of the model, including a “smart search engine” with references to scientific sources, a relationship graph for identifying hidden patterns and mechanisms for working with incomplete or ambiguous information. Eventually, these capabilities are expected to merge into a multi-agent system capable of solving complex analytical tasks automatically.
The main domestic applications include scientific analytics, decision support for evaluating emerging research fields and scientific schools, patent risk analysis, competitor monitoring and assistance for researchers working with large datasets.
The broader concept of specialized national and industry-focused large language models is also gaining traction internationally. The Russian project could attract attention as an example of adapting AI systems to the scientific terminology of a non-English-speaking country and as an approach to reducing the resource intensity of large models.
Direct exports of the model itself may be limited by its language specialization. However, the methodology for building sector-specific language models on top of national data corpora could prove valuable for countries facing the same imbalance created by the dominance of English-language data in AI systems. The most promising markets include CIS countries, BRICS members and other friendly jurisdictions.
Professional AI Instead of a Universal Chatbot
The significance of the HSE University project lies not only in the neural network itself, but in what it represents: a transition to a new stage in Russian AI development, from universal chatbots to specialized professional models. The main practical benefit is higher quality and faster processing of Russian-language scientific and technical information.
The system’s effectiveness will ultimately depend on the completeness of its data corpus, the transparency of source references and protection against hallucinations. If the project reaches industrial-scale deployment, it could occupy a niche as a specialized AI assistant for Russian science, technological expertise and innovation policy. In practice, this is one of the building blocks of a future “scientific AI” infrastructure in Russia – a professional tool for researchers, analysts and engineers.
