Yandex Introduces Sona, the World's First AI Model to Replace a Full Recommendation Pipeline Without Hand-Engineered Features

Yandex has introduced Sona, a generative AI model designed to handle the full recommendation process within a single model, from selecting relevant items to ranking the final results. In a seven-day live production experiment on Yandex smart speakers, Sona replaced the existing recommendation system for the test group and increased listening time for recommended tracks by 6.3%. The results demonstrate a new approach to building recommendation systems: replacing a complex multi-stage stack with a single model while improving recommendation quality.

Sona is the first publicly documented recommendation system to demonstrate in live production that a single model can replace a full multi-stage recommendation cascade without hand-engineered features while delivering statistically significant gains in user metrics.

How Sona works

In the experiment, Sona replaced a production recommendation pipeline of dozens of algorithms, from candidate generation through separate pre-ranking and ranking stages. Bringing these stages into a single model reduces the complexity of maintaining and optimizing multiple interconnected systems and can make the recommendation stack easier to develop and iterate.

Sona also operates without hand-engineered features. Traditional recommendation systems can rely on hundreds of features designed and tuned by engineers to describe users, content and their interactions. Sona instead learns directly from user interaction data, reducing the manual work required to build and maintain these signals.

The model analyzes a person's interaction history — including what they choose, skip or return to — and represents recommended items using learned Semantic IDs that capture information about both the content itself and how users interact with it. Its architecture combines an encoder, an autoregressive decoder and a ranking module to generate and rank recommendations within the same system.

Production results

Yandex tested the model for seven days in music recommendations delivered through smart speakers with Alice AI. During the experiment, Sona replaced the existing recommendation stack for the test group. Listening time for recommended tracks increased by 6.3%, requests to replay a track rose by nearly 18%, and the number of active listeners increased by 4.5%. Likes also rose by 11.4%, while the number of highly engaged listeners increased by 7.4%. The lift in active listeners was 2.4 times the gain delivered by the previous recommendation-system update, which had itself sharply improved recommendation quality.

Industry context

Sona is part of a broader shift in recommendation technology toward larger, unified models capable of performing tasks previously handled by separate components. Kuaishou's OneRec has replaced a cascaded recommendation framework with a unified generative model and delivered gains in live production, while Pinterest's UniPinRec has combined retrieval and ranking within a shared production architecture.

The shift matters because recommendation systems underpin a wide range of digital products, from music and video discovery to e-commerce, social feeds and advertising. Improvements in recommendation quality can influence engagement and retention as well as commercial metrics such as conversion.

Sona builds on more than a decade of Yandex research in recommendation systems. The company has previously introduced transformer-based recommendation technologies including ARGUS and Gryphon. In 2025, Yandex also open-sourced Yambda, one of the largest publicly available datasets for recommendation-system research, containing nearly 5 billion interactions and already used in research by teams at Kuaishou and Meta. Researchers at ByteDance have also cited Sona in their own recent work on unifying the recommendation cascade. The Sona team has published the model's development and production experiment results in a research paper available on arXiv.