Yandex Introduces Sona: A Generative Recommender That Replaces the Full Recommendation Cascade

Yandex has unveiled Sona, a single generative model designed to replace the multi-stage recommendation cascade — typically composed of retrieval, ranking, and re-ranking components — with one unified system. Sona generates recommendations end-to-end, eliminating the error accumulation and optimization misalignment that occurs when separate models hand off between pipeline stages. For developers building recommendation systems, this approach represents a significant architectural simplification that could reduce infrastructure complexity while improving relevance coherence. Yandex's scale as a search and content platform makes Sona a credible proof point that generative recommendation is viable in high-traffic production environments. Teams maintaining multi-stage recommendation pipelines should watch this space closely as generative alternatives mature.