Russia Enacts Framework for Large AI Models

The law creates two classifications for large foundation models. Izvestia reports that a sovereign large foundation model must be fully developed by a Russian legal entity, monitored throughout its lifecycle, technically reproducible, and hosted in Russian data centers.

A national large foundation model, according to Izvestia, must also be developed by a Russian legal entity but may incorporate foreign components distributed under open licenses. The Kremlin states that the law sets requirements for both sovereign and national models.

Izvestia reports that qualifying models contain at least 1 billion parameters and can perform intellectual tasks at a level comparable to or exceeding humans, while serving as a basis for multiple software products. ForkLog separately reports that the regulations apply to models with at least 1 billion parameters and services with up to 500,000 daily users; the available reporting does not clarify how that user threshold interacts with the broader statutory framework.

According to Izvestia, developers of qualifying models can receive state support, including financial, property, guarantee, and information measures. The publication also reports that developers can train models using data from federal and regional information systems, with the government to determine access procedures in coordination with security authorities.

The Kremlin's announcement specifies that the law assigns powers across the presidency, government, federal executive bodies, other state institutions, state corporations, and the Central Bank. United24 Media reports that the president is responsible for approving Russia's national AI-development strategy, while the government oversees the development and deployment of large foundation models and determines circumstances in which sovereign or national models are required.

For ML teams, legal access to state-held datasets can materially affect the availability of Russian-language, public-administration, and domain-specific training corpora. Comparable state-supported model programs also tend to make data-governance controls, provenance records, infrastructure location, and reproducibility more consequential design constraints.

The Kremlin states that the law introduces requirements for labeling content generated with large foundation models. United24 Media characterizes the enacted provision more narrowly: platforms with more than 500,000 daily users must provide technical means for labeling AI-generated content, while developers and platforms can choose the labeling format.

United24 Media also reports that service owners must inform users about ownership and terms of use for AI-generated content. The outlet writes that the law treats training on copyrighted material as non-infringing when the data was legally obtained and technical protections were not bypassed.

Izvestia and United24 Media report that sovereign and national models are required to comply with Russian law and account for "traditional Russian spiritual and moral values." This combines model-development support with content and governance conditions that are explicit in the foundation-model legislation. Across jurisdictions, such requirements can add evaluation criteria beyond conventional quality, safety, and copyright testing.

ForkLog and United24 Media report that most provisions take effect on September 1, 2026. The practical effect will depend on subsequent government rules governing model designation, dataset access, support eligibility, platform labeling mechanisms, and enforcement.