Fifteen independent heuristic candidate generators and a heavy ranking layer powered by hundreds of handcrafted features have served as the unspoken standard in recommender systems for years. However, this architecture came at the cost of spiraling technical debt and fragmented pipelines that drained engineering resources.
The Yandex Music team has published a technical report (arXiv: 2608.11015) outlining the result of a year-long engineering effort: the Sona model. The developers dismantled the multi-stage cascade, replacing it with a single end-to-end trained transformer.
Architectural trade-offs and distillation
Consolidating a recommendation pipeline into a single neural network sounds logical, but in practice, it runs into fundamental inference constraints. Classical architectures like OneRec rely on a separate ranking model fed with manual features. To build a truly monolithic system, candidate generation and final scoring must rely on a shared user embedding encoder that runs exactly once per request.
The core challenge lies in data sparsity: users hit 'like' and loop tracks too infrequently to train a model on fine-grained ranking directly. Yandex engineers solved this using knowledge distillation. They first trained a Teacher Ranker on a full year of logs—a heavy offline model that never touches production, but instead passes its scores to a lightweight ranking module inside Sona.
Catalog tokenization and latency management
Architecturally, the model is built as an encoder-decoder trained chronologically on next-token prediction, where each track is encoded via a Semantic ID. Along the way, the team hit a paradox: theoretically, scaling the token vocabulary should yield higher precision, but in practice, the model failed to navigate the expanded space, causing metrics to drop sharply.
Sona is a single model. Previously, the Yandex Music production pipeline included 15+ candidate generators alongside ranking built on hundreds of handcrafted features. All of that was replaced by a single, end-to-end trained model.
Autoregressive generation over long user histories remains notoriously compute-heavy. To meet strict streaming latency SLAs, the team implemented History Compression, custom CUDA kernels, and an optimized radix top-k algorithm under the hood.
In Yandex Music's production environment, the model proved the viability of the approach: A/B testing showed a 4.53% increase in active users, a 6.30% jump in listening time, and an 11.42% boost in likes on top of an already highly optimized baseline. The unit economics pencil out by eliminating the maintenance overhead of a microservice zoo, though it requires expensive online transformer training and a complete overhaul of the ML infrastructure.