Russian LLMs Take the Top 6 Spots in the SLAVA Benchmark, but They Need Global Data
In early 2026, Russian LLMs took the top 6 spots in SLAVA, a test of 14,000 questions on Russian history and geography. In November 2024, American and Chinese models led the ranking. But the Russian-language internet is 10–20 times smaller than the global one, and ITMO estimates that Russian texts should make up only 5–30% of training data. Without international data, models will fall behind foreign counterparts, so choose a model by testing it on your own tasks.
AI-processed from Habr AI; edited by Hamidun News
Imagine you are building an assistant for lawyers. You ask a neural network about Russian law, then ask it to break down a fresh English-language scientific paper. A model that has read a lot of Russian text will handle the first question better.
A model trained on a global corpus will handle the second. That is the subject of a study by ITMO's AI Institute, prepared with experts from MSU, RANEPA, HSE, AIRI, Yandex, Sber, T-Bank and VK. In early 2026, Russian LLMs (large language models) took the top 6 spots in the SLAVA ranking, a test of knowledge of Russian history, law and culture.
But to keep that advantage, the models need broad access to data, both Russian and international. If you are choosing a model for a product, this bears directly on your decision.
What the SLAVA test showed
SLAVA tests models on 14,000 questions about Russian history, social studies, sociology and geography. In early 2026, domestic LLMs held the first six places in the ranking. As recently as November 2024, American and Chinese models made up the top three. The picture changed in less than a year and a half.
The authors attribute this not only to developer investment but also to a growing share of Russian data in training sets. A model that has read more Russian text does not just speak Russian better. It knows Russian realities, law and cultural context better. For tasks with a Russian focus, such as advising on local regulations or working with government documents, a local model can offer a noticeable advantage.
Why Russian alone is not enough
The Russian-language internet is roughly 10–20 times smaller than the global one. Modern models are pretrained on corpora of trillions of tokens: web pages, books, scientific publications, open-source code and encyclopedias. According to the researchers, there is enough Russian data to make up part of a multilingual corpus, but not enough for a competitive general-purpose model trained on Russian alone.
The experts put the optimal share of Russian-language data in a training corpus at 5–30%. Other languages should supply the rest, since they contain different useful information, and the English-language corpus remains the key carrier of scientific knowledge. A model isolated from international data will fall further behind foreign counterparts and perform worse on tasks outside the local language context.
Russian teams also have it harder than foreign ones. They depend more heavily on open sources because they have fewer resources to collect, buy and digitize data themselves. Any restriction on the available corpus hits them harder.
What the authors propose
First, digitize data that is not yet in the open internet: library and archive materials, and educational, scientific, cultural and professional collections. This could improve models in medicine, law, construction, geology and other fields where Russian knowledge and standards matter.
Second, stop duplicating work. Alexander Bukhanovsky, director of ITMO's AI Institute, notes that every large company currently builds models for itself and keeps its developments closed. Instead of each company creating an agent that reads the same professional literature, they could exchange such agents or offer them on a commercial basis. Companies could then assemble systems for their own tasks from ready-made components.
Third, regulate copyright carefully: the rules should protect rights holders without shrinking the pool of information available for training.
How to read this in practice
This is an industry study, not an independent experiment. Its authors are market participants and research organizations with a stake in data access. The SLAVA benchmark measures knowledge of Russian context, not a model's general intelligence or the quality of its code.
The practical takeaway for developers and product teams is simple. Do not choose a model based on a single ranking. Test it on your own tasks: the local advantage is real, but it shows up where the Russian language and Russian realities matter. Closed corpora such as regulatory databases and professional literature can become your edge when fine-tuning or building document search. And the market as a whole, in the authors' view, has a chance to become a full-fledged infrastructure layer for Russia if it takes the global technology base, strengthens it with domestic data and moves products quickly into deployment.
- Nov 19, 2026Open DaisyRose-27B solves 12% of Terminal-Bench tasks, up from 5.6%
- Nov 19, 2026The same Gemma 4 wrote a backend that passed 80% of checks in opencode, but only 41% in OpenHands
- Nov 19, 2026OpenAI's coding agents would cost $2400 a day per researcher at API prices
- Nov 19, 2026AWS AgentCore Agents Can Now Reach a Private Keycloak Behind a Corporate CA
Key Questions Answered
What the SLAVA test showed?
SLAVA tests models on 14,000 questions about Russian history, social studies, sociology and geography. In early 2026, domestic LLMs held the first six places in the ranking. As recently as November 2024, American and Chinese models made up the top three. The picture changed in less than a year and a half. The authors attribute this not only to developer investment but also to a growing share of Russian data in training sets. A model that has read more Russian text does not just speak Russian better.
Why Russian alone is not enough?
The Russian-language internet is roughly 10–20 times smaller than the global one. Modern models are pretrained on corpora of trillions of tokens: web pages, books, scientific publications, open-source code and encyclopedias. According to the researchers, there is enough Russian data to make up part of a multilingual corpus, but not enough for a competitive general-purpose model trained on Russian alone. The experts put the optimal share of Russian-language data in a training corpus at 5–30%.
What the authors propose?
First, digitize data that is not yet in the open internet: library and archive materials, and educational, scientific, cultural and professional collections. This could improve models in medicine, law, construction, geology and other fields where Russian knowledge and standards matter. Second, stop duplicating work. Alexander Bukhanovsky, director of ITMO's AI Institute, notes that every large company currently builds models for itself and keeps its developments closed.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.