Yandex builds its own AI model as Russia pushes sovereignty
The Russian tech giant's new language model packs 80 billion parameters but only activates 3 billion at a time, a design trick that makes it far cheaper to run than its size suggests.
Yandex has developed an 80-billion-parameter large language model from scratch called Alice AI Foundation, the company told Vedomosti. The Russian tech giant said the model does not rely on third-party components.
The company has published the model’s weights under a license allowing commercial and research use, opening it to developers who want to run or further train the model.
Alice AI Foundation was trained on 18 trillion tokens and has 3 billion active parameters, according to Yandex. The company plans to use the model as the foundation for a future reasoning system and the agent-based features of Alice AI.
The move comes as Russia develops a legal framework for sovereign AI models. The law distinguishes sovereign models from national models, requiring the former to use domestic components including software, architecture and infrastructure. However, the formal process for assigning sovereign status will not begin until March 2027.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Russian AI experts said the release reinforces Yandex’s role as one of the country’s major developers of fundamental models. At the same time, they noted that Alice AI Foundation is a base model rather than a finished reasoning system and is smaller than leading models with hundreds of billions or trillions of parameters.