Yandex introduces Sona, a generative recommender that replaces recommendation cascade

In the test, Sona outperformed the existing pipeline, delivering +4.53% increase in active users, +6.30% total listening time, +11.42% likes, +17.99% repeat commands, and +7.37% deeply engaged users. The architecture eliminates hand‑engineered features, relying on logged event fields and learned Semantic IDs, and uses a 0.6 B‑parameter teacher for distillation. The system processes 8,192 past events with a hybrid attention scheme, achieving about half the inference cost of full attention while maintaining quality.

While Sona is not publicly available and cannot be deployed outside Yandex, the results suggest that end‑to‑end generative recommenders can deliver stronger performance than traditional cascades. The approach may influence future recommender designs across the industry.