Yandex launches sona: a unified generative recommender system

Yandex has introduced Sona, a generative AI model that streamers the traditional multi-stage recommendation pipeline by integrating candidate generation and ranking into a single system. The model was tested in a seven-day live production experiment on Yandex smart speakers, where it replaced over 15 candidate generators, the pre-ranking stage, and the ranking stage with a single transformer. Sona uses logged event fields and learned Semantic IDs, eliminating the need for hand-engineered features. The system employs a frozen multimodal LLM to generate Semantic ID tuples, which are then scored by a Ranking Module. Training involved a 0.6B-parameter transformer trained on a year of engagement events, with a joint loss function combining next-item-prediction pre-training and multi-head ranking fine-tuning. Sona’s performance showed a 2.35x uplift in Active Users compared to previous systems. While not the first end-to-end generative recommender, Sona combines a full cascade replacement, no hand-engineered features, and a distilled ranker, validated online. Yandex’s Sona is distinct from OneRec, which uses engineered user features and reinforcement learning, and from Meta’s HSTU, which reframed recommendation as sequential transduction in 2024.