National language model: ambitions, reality, and the cost of sovereignty

Discussions about the need for Russia's own large language model have been ongoing for years, but they remain largely confined to declarations and isolated initiatives. A fresh post on Habr by a practicing specialist forces a sober look at the problem—without patriotic rhetoric and without technological pessimism. And the picture that emerges proves far more complicated than government officials and corporate strategists would like.

The thesis itself is straightforward: creating a competitive LLM from scratch is not a project but an ecosystem. Three foundational pillars—talent, hardware, and institutional knowledge—sound obvious, but the devil lurks in the details of each. Let's start with talent.

We're not talking simply about programmers who know Python and are familiar with transformer architecture. What's needed are deep mathematicians capable of working at the frontier of optimization theory, specialists in distributed computing, and engineers who understand the nuances of training models with hundreds of billions of parameters. Such people number in the mere thousands worldwide, and most are concentrated in the ecosystems of Google, Meta, OpenAI, and a handful of Chinese tech giants.

The Russian market competes for these specialists under decidedly unfavorable conditions—and it's not just about salaries, but also access to world-class computing infrastructure.

The situation with hardware is even more acute. Training modern LLMs at the scale of GPT-4 or Claude requires clusters of thousands of graphics accelerators in the NVIDIA H100 class or their equivalents. Sanctions restrictions significantly impede legal supplies of top-tier chips to Russia, and domestic alternatives with comparable performance do not yet exist. Projects like Elbrus and Baikal address different challenges and lag behind leaders by generations, not years, in computational power. Parallel imports and workaround schemes might cover specific needs, but building systematic world-class model training on them is a fantasy.

However, the author of the post rightly points to the most underestimated factor—the presence of institutional knowledge. This concept is broader than simply accumulated experience. It's a culture of engineering solutions that passes from project to project, team to team.

It's institutional memory of thousands of experiments, failed approaches, and non-obvious discoveries that cannot be extracted from academic papers. OpenAI traveled the path from GPT to GPT-4 in five years of continuous iteration. Google DeepMind accumulated expertise over more than a decade.

Attempting to skip this phase through "effective management" and budget injections is a typical mistake, which the author delicately yet precisely describes with the phrase that "the mere presence of desire and money does not always lead to the desired result."

It's important to understand the context: Russia is not starting from scratch. Yandex has the YandexGPT family, Sber is developing GigaChat, and other initiatives exist. But the gap between these products and world leaders remains significant, and it risks not narrowing but growing—the pace of frontier model development has only accelerated in the past two years. China, possessed of incomparably greater resources and its own chip manufacturing, has still not managed to confidently catch up to American leaders, though it has substantially closed the gap thanks to the DeepSeek model and several other breakthroughs.

For the industry, this discussion has quite practical implications. If the bet is on fully sovereign development, it means years of investment with no guaranteed result. The alternative path is developing expertise in fine-tuning and adapting open models like Llama or Mistral to the specific needs of the Russian-speaking market. This approach is more pragmatic, cheaper, and delivers results faster, though it doesn't solve the problem of strategic dependence.

Ultimately, the question of a national LLM is not a technical but a political-economic one. Is the state ready to invest not in showcase projects but in fundamental infrastructure: education, research centers, access to computing? Is business ready to think in horizons of ten to fifteen years rather than quarterly reports? Until the answers to these questions become clear, the conversation about a world-class sovereign language model remains more an exercise in strategic thinking than a roadmap.