Yandex has introduced Sona, a generative AI model designed to function as a unified platform for generating and ranking user recommendations. The model replaces traditional multi-algorithm systems by directly analyzing user behavior to suggest relevant content, eliminating the need for manual feature selection.
The company has already integrated Sona into the music recommendation system for its smart speakers powered by Alice. Following a one-week deployment, the model increased the listening time for recommended tracks by 6.3 percent and boosted repeat play requests by nearly 18 percent.
The implementation of Sona also resulted in a 4.5 percent growth in the active listener base, which the company reports is 2.4 times higher than the growth observed during the previous system update. Sona builds upon Yandex's earlier developments, including the ARGUS transformer-based system from 2025 and the Gryphon architecture introduced in 2026.
Yandex continues to contribute to the field of recommendation systems, having released the Yambda dataset in 2025, which has been utilized by researchers from Kuaishou and Meta. The Sona model has received positive feedback from ByteDance researchers involved in the TikTok recommendation system.
