Sona model boosts Yandex Music engagement by over 4% in A/B test

Meta: Meta is a social technology company operating platforms with extensive recommendation needs for feeds and content discovery. Its research has introduced paradigms like index-as-model approaches that unify retrieval components under a single neural architecture. Meta researchers have also engaged with open datasets in the recommendation space, consistent with the collaborative trends mentioned in the news. Sona: Sona is a generative recommender model developed by Yandex that unifies candidate generation and ranking in a single architecture using shared user representations from engagement sequences. It operates without hand-engineered features and was trained jointly with objectives for next-token prediction and distillation. The model was deployed in an online A/B test on Yandex Music, where it replaced a multi-stage cascade and demonstrated improvements in user engagement metrics. Spotify: Spotify is a leading audio streaming platform that invests in research on large language models and unified systems for personalization, search, and discovery. Its teams have developed frameworks like NEO that adapt LLMs into steerable models combining multiple recommendation capabilities. This reflects the same industry-wide push toward unified recommenders referenced alongside the Sona release. Tencent: Tencent is a major technology conglomerate with significant interests in social media, gaming, and digital entertainment platforms that use advanced recommendation systems. Researchers associated with the company have explored unified blocks for scalable recommendation models that handle feature interactions and sequential behaviors. Their work supports the movement toward simplified, single-model pipelines noted in recent industry developments. Kuaishou: Kuaishou is a short-video platform operator that develops and deploys large-scale recommendation models for user engagement. Its teams have published on unified model-centric scaling frameworks that improve efficiency across multiple product surfaces. This work contributes to the industry movement toward simplified, generative recommendation systems. ByteDance: ByteDance is a global technology company known for its short-video and social platforms that rely heavily on sophisticated recommendation algorithms. Its research teams have contributed to the industry shift toward unified recommendation models through publications exploring multi-distribution learning and transformer-based architectures. This aligns with the broader trend highlighted in the Sona announcement. Pinterest: Pinterest is a visual discovery and social platform that deploys large-scale recommendation systems for content ranking and retrieval. Its engineering teams have advanced full-stack unification of generative retrieval and ranking in production environments. Such efforts exemplify the broader shift by major platforms toward consolidated recommendation architectures. Yandex Music: Yandex Music is a music streaming service operated by the Russian technology company Yandex, offering personalized recommendations across devices including smart speakers with Alice. It has invested in advancing recommendation systems through internal research projects such as ARGUS and Gryphon. Sona was first tested and validated on one of its largest recommendation surfaces, My Vibe on smart speakers. Industry Trend: Major technology platforms are actively publishing research on unified generative and transformer-based recommendation architectures to replace traditional multi-stage systems. Research Collaboration: Open datasets from recommendation research have been cited and built upon by teams at leading global technology companies in recent academic work.