Yandex has unveiled Sona, a generative AI model designed to replace complex cascades of recommendation algorithms. Unlike traditional systems that require manual feature selection and multiple stages of processing, Sona performs both selection and ranking of content independently by analyzing user behavior, including selections, rejections, and revisits.
The company first tested Sona in the music recommendation system for smart speakers with Alice. During a one-week trial, the model replaced dozens of existing algorithms. The experiment resulted in a 6.3% increase in listening time and an 18% rise in track repeat requests within the test group. The number of active listeners grew by 4.5%, which is 2.4 times higher than the previous update.
The development of Sona involved engineers from Alice, Yandex Music, and the company's R&D team. The findings were published in a research paper on the international platform arxiv.org. This model follows the company's previous developments, including the ARGUS system introduced in 2025 and the Gryphon architecture implemented in 2026.
Yandex has been developing recommendation technologies for over a decade. In 2025, the company released Yambda, a large-scale open-source dataset for recommendation research, which has been utilized by scientists from Kuaishou. Additionally, researchers from ByteDance have acknowledged the performance of the Sona model.
