Yandex replaces dozens of recommendation algorithms with a single model, Sona: +6.3% listening time

Anyone who has built a recommendation system knows this pain. First, one model selects candidates from millions of items, then another ranks them, a third adds diversity, and a fourth accounts for context. Each is trained separately, each has its own metrics, and maintaining this sprawl gets more expensive all the time. Yandex has shown that the chain can be replaced with a single model.

What was built

The R&D team, together with engineers from Alice and Yandex Music, developed Sona, a generative neural network. It looks at the user's actions and decides on its own what to recommend and in what order. The architecture is proprietary, and training was done from scratch on Yandex data, without fine-tuning anyone else's weights.

In an A/B test on Alice-powered smart speakers, Sona replaced a system of dozens of algorithms. Compared with the control group, total listening time rose by 6.30% and the number of likes by 11.42%. These are notable figures for a service where recommendations have been hand-optimized for years.

How it works, with a simple example

Imagine a user saying to the speaker in the evening, "play something." A classic system first narrows the catalog down to hundreds of candidates, then sorts them with different models. Sona works differently: it generates a recommendation as a sequence, as if writing a playlist track by track. Each next track takes into account the previous ones and the listener's reaction.

To avoid going through the entire catalog of millions of tracks, the model finds the relevant items directly. This is critical for speed: brute-force search at this scale would be too expensive.

There is also a "teacher." According to the authors, it is used to guide the model's training. Yandex gives details in the full article on Habr, and those details deserve a closer look separately. Another important part is continual training on listeners' fresh actions. Tastes change quickly, and the model should not live on yesterday's data.

What this means in practice

For product managers, the main takeaway is simplicity. One model pipeline instead of dozens means fewer points of failure, unified metrics, and easier experiments. The team does not synchronize updates across five models but improves just one.

For developers, this is a trend already visible among the big players: generative approaches are displacing multi-stage recommendation pipelines. If you maintain your own recommendation system, it is worth considering which stages could be merged.

Limitations

The authors themselves are cautious: much remains to be verified before broad deployment. The test covered a single scenario, music on Alice-powered speakers. It is unknown how Sona would behave in video, e-commerce, or search, where the goals are different. There is also no data yet on inference and training costs compared with the old system. For small teams without Yandex-scale data, repeating this approach will be hard: training from scratch requires both data and compute.

Bottom line: this is not a revolution for every case, but a strong signal. If a large recommendation system benefits from replacing a chain of algorithms with one model, then the question of consolidation is worth asking of your own stack too.