Yandex Music replaces its recommendation system with one transformer model

Yandex has tested a single model that replaces the multi-stage system behind Yandex Music’s song recommendations, according to a report from MarkTechPost. The model, called Sona, is a transformer, a neural network architecture built to process sequences of data, such as a user’s listening history. In a live A/B test, a side-by-side experiment comparing the new system against the old one, Sona lifted user likes by 11.42%.

Most large-scale recommendation systems run as a cascade: a pipeline of separate models, each handling one stage. One model generates a shortlist of candidate songs from a catalog of millions. Another ranks that shortlist to decide what the user sees first. These stages typically depend on hand-engineered features, manually crafted inputs that data scientists build to describe a user or an item, such as how often someone skips a genre. Sona, according to the report, performs both candidate generation and ranking in one pass, without those manually built inputs.

What the result suggests

The reported gain points to a broader shift that AI labs and platforms have been testing: folding several specialized models into one general-purpose model that learns directly from raw interaction data. If a single transformer can match or beat a cascade built from multiple purpose-built models, it reduces the number of systems an engineering team has to maintain, tune and keep in sync. That is the kind of consolidation other companies with large recommendation pipelines, such as video or e-commerce platforms, are likely watching closely.

The source material does not say how large Sona is, what data it was trained on, or how long the A/B test ran. It does not say whether the 11.42% lift in likes translated into other outcomes Yandex Music might track, such as listening time or skip rates. It is also not stated whether Sona has been rolled out to all Yandex Music users or remains confined to the tested group. MarkTechPost’s report describes the model’s role in the test but does not include those details, so they should be treated as open questions rather than assumed.

It is also unclear from the available material whether Sona was built entirely in-house or draws on techniques Yandex has published before, and whether the company plans to release a technical paper or open-source any part of the system. Yandex has previously open-sourced other engineering work, but nothing in this report confirms that plan for Sona specifically.

What to watch next is whether Yandex publishes further detail on Sona, such as model size, training setup or a technical paper, and whether the company extends the single-model approach beyond Yandex Music to its other recommendation products.