Field noteNov 19, 20265 min read1 source

ITMO has published a study of Russian large language models, their training data and evaluation landscape

In one minute

ITMO has published a study of Russian large language models, their training data and evaluation landscape, combining an overview of models and benchmarks with 14 expert interviews. Its central argument is that competitiveness depends not just on compute or data volume, but also on curation, access to usable local datasets, multilingual training and the institutions needed to support development.

What changed: ITMO published a study of the Russian LLM market, covering 16 models, 12 benchmarks, seven training stages and 14 expert interviews.

What happened

ITMO’s open-source ecosystem initiative describes the publication as an analysis of the Russian LLM market and the data used to train models. The study reviews 16 domestic models, 12 Russian-language benchmarks and seven stages of training, ranging from tokenizer adaptation to verifiable rewards. ITMO says it conducted 14 in-depth, semi-structured interviews with people from Russian technology companies, independent teams, academia and government organizations. The page names participants from organizations including Yandex, Sber, T-Bank, VK, AIRI, Moscow State University and ITMO, as well as independent developers.

The study presents six conclusions. First, it argues that quality improvements depend on rules, data, infrastructure and people, not on any one input. It says Russia has taken an initial legal step through federal legislation on using publicly available content for model training, while further progress requires computing resources, business deployment, research methods, benchmarks and specialist education. The page does not lay out the law’s provisions in detail; it points readers to the full report for its argument and expert quotations.

Second, ITMO identifies existing but underused collections—library, archive, educational, cultural and professional data—as a potential growth source. The study says organizational, legal and technical barriers keep such material from models, and argues that opening and digitizing collections that contain no sensitive information and can be made public could benefit domestic models. Third, it points to Russian-language and culturally specific tasks as areas of potential advantage. It reports that domestic LLMs held the top six positions in the SLAVA benchmark at the beginning of 2026, and names medicine, law, construction and geology as fields where local context may matter.

The remaining conclusions stress that Russian models should remain multilingual; that carefully curated smaller corpora can compete with much larger raw datasets, with DeepSeek cited as an example; and that progress requires coordinated work among companies, government, universities and open-source communities. ITMO argues that non-Russian data contributes to a model’s understanding of the world, not only its vocabulary. Its market dashboard provides context figures drawn from external sources, including an estimate that the compute gap between the United States and Russia is about 660 times. The report is downloadable as a PDF from the ITMO page.

What the archive adds

We have not covered this before; it opens a new thread.

Trend signal

This does not yet fit a pattern New Runtime tracks.

Why now

The study situates its analysis in 2026, amid active releases of Russian open models, benchmarking and debate over rules for training data. ITMO frames the report as a view of the market’s current constraints and opportunities, rather than as a response to one particular model release.

Why it matters

For teams building with language models, the report’s most useful implication is that data strategy is not reducible to collecting more text. Curation, provenance, lawful access and task-relevant evaluation shape whether a corpus can improve a model in a particular language or domain. The study’s focus on archives, libraries and professional collections also highlights a practical bottleneck: valuable data may exist but remain unusable because it has not been digitized or because access conditions are unresolved.

This is particularly relevant when a team is considering domain adaptation, retrieval corpus construction or an internal evaluation program. A collection may be rich in subject knowledge while still being unsuitable for training or hard to use operationally: rights may be unclear, records may be poorly digitized, and its format or coverage may not match the target task. The study does not turn those constraints into an implementation recipe, but it makes them visible as engineering and organizational work rather than as background paperwork.

The report is about the Russian market, but the underlying questions travel: which collections are fit for training, which are legally and technically accessible, and what benchmarks reflect the work users actually need done? Its findings are a reason to examine those questions in a model program, not proof that any particular data investment will yield better performance. Teams should treat the reported benchmark standing and interview-based conclusions as context to investigate against their own workloads.

Limits and uncertainty

The public page summarizes the study’s scope and conclusions but does not reproduce the full evidence, interview protocol, benchmark test conditions or model-by-model comparisons. The reported SLAVA ranking is a time-specific result from the beginning of 2026; it does not establish performance across other benchmarks, languages or production tasks. ITMO’s statements about data quality and curated corpora are research conclusions, not a controlled comparison detailed on the page. Several market figures on the dashboard are attributed to outside organizations, so their definitions and methods should be checked at those sources. The page links to a full PDF, but the claims here should not be read as independently verified measurements of the Russian compute gap or market outlook.

Reader takeaway

ITMO’s study frames Russian LLM competitiveness as an ecosystem problem: useful data must be accessible and carefully prepared, while models need multilingual coverage, relevant benchmarks, compute and skilled teams. The headline is not simply that local models lead on a benchmark; it is that sustained progress depends on making data and evaluation infrastructure usable.

Evidence / sources

  1. 1opensource.itmo.ru/llmresearch
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicModelsExplore the models topic hub.
  2. 02archiveField NotesOpen the latest editorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract