MOSCOW, RUSSIA / RankWire.AI / – Russia’s Federation Council approved a framework bill on July 17 that sets national rules for developing and using large artificial intelligence foundation models. The measure creates legal definitions, assigns regulatory powers and establishes support for qualifying domestic developers. It also sets requirements for model control, data storage, user notices and AI-generated content. The bill cleared the State Duma on July 8 and now requires presidential approval before taking effect.
Under the bill, a large foundation model is software that performs many intellectual tasks at a level comparable with human ability. The definition covers systems that provide information, make decisions or forecast results based on goals set by a person. A qualifying model must contain at least 1 billion parameters. The framework also sets principles covering technological sovereignty, human rights, personal choice, security and compliance with Russian law.
The legislation creates separate sovereign and national model categories. A sovereign model must come from a Russian legal entity and remain under domestic control. It must use data centers inside Russia. Developers must preserve the technical ability to reproduce the full development cycle, including training and original parameters. A national model follows similar rules for ownership and data processing. However, it may include foreign components distributed under open licenses.
Domestic models gain formal status
The government may set support measures for companies that build, deploy or operate qualifying foundation models. Those measures can include access to state-held data needed for training. The law also allows authorities to define settings where only sovereign or national models may operate. That power covers government information systems and other sensitive areas. Separate laws and presidential decrees may set rules for defense, security, public order and property protection.
Operators of websites, applications and social platforms face a new content rule. Services with more than 500,000 daily users must provide a tool for marking AI-generated audio or visual material. The rule gives users a way to add an information notice to qualifying content. Developers and users will determine the label format through service agreements. Platforms do not have to label every item automatically, but they must offer the technical option.
Copyright and implementation rules take shape
AI service providers must tell users who owns rights to generated material. They must also explain access terms and whether users may transfer the content. The bill separately addresses copyrighted works used for machine learning. It treats analysis for extraction, comparison, classification and pattern detection as lawful when developers obtained access legally. Training may use protected works when no one bypassed technical restrictions that limited access to the material.
Most provisions are scheduled to start on September 1, 2026, if the president signs and publishes the measure. Rules for domestic model status, developer duties, content marking and intellectual property will start on March 1, 2027. Existing systems receive a transition period until September 1, 2032, when their data processing and storage remain inside Russia. Until presidential approval and official publication occur, the measure remains an approved bill rather than an enacted federal law.