The Russian ML automation market is shifting from monolithic models to specialized agentic systems. Businesses are moving away from using a single large model for all tasks, such as contact centers or legal analysis, in favor of ensembles of smaller, task-specific models ranging from 1B to 80B parameters.
Data security and legal compliance remain critical drivers for this transition. Under Russian law, specifically the 152-FZ regulation, using foreign cloud APIs for processing non-anonymized personal data poses significant risks. Companies are increasingly opting for local infrastructure or Russian cloud providers to ensure compliance with data localization and processing requirements.
The cost of deploying large open-weight models, such as the Qwen3-235B, is substantial, with hardware clusters often exceeding $200,000. Despite high costs, these models require fine-tuning to achieve the precision needed for specific corporate tasks, as general-purpose models often lack the necessary domain-specific accuracy.
Industry trends indicate that RAG and fine-tuning serve complementary roles: RAG acts as a reference for fact retrieval, while fine-tuning helps models master specific formats and styles. Consequently, companies are building separate microservices for different departments, such as HR, legal, and accounting, each with its own data sources and access policies.
