Russia revises AI regulation bill: sovereign models remain but can now be trained on non-Russian data

The Russian government has circulated a revised version of its artificial intelligence regulation bill, according to RBC, Izvestia, and Kommersant. The document has been shortened significantly, from 21 articles to 13, and renamed from "On the Fundamentals of State Regulation in the Areas of Application of AI Technologies" to "On Supporting the Development of AI Technologies."

The new version no longer addresses AI broadly; instead, it focuses on the development, deployment, and use of large foundation AI models. It drops the concept of a "trusted" AI model, which was previously planned for systems used in government and critical information infrastructure. Only "sovereign" and "national" models remain, a status granted to models developed by Russian legal entities that process user requests and store data in Russia, and comply with Russian law and moral and spiritual values.

A source familiar with the IT market told RBC that Sberbank's model meets the criteria for a sovereign large foundation model, while Yandex's model qualifies as a national one.

The revised bill also removes the requirement that models seeking "sovereign" or "national" status be trained exclusively on data of Russian origin or developed solely by Russian citizens. Izvestia reports that the new version contains no provisions allowing the government to ban foreign neural networks in Russia. Deputy Prime Minister Dmitry Grigorneko said businesses would retain the right to independently choose the technological solutions that best serve their needs.

Experts noted that many Russian AI services are built on open-source foreign models, such as DeepSeek, and a ban would have shut down working products. The first version of the bill was criticized by Russian business representatives, who argued it would raise costs, slow product launches, and push development to other jurisdictions.