Yandex has released its Alice AI Search Pretrain model on the Hugging Face platform. This base model, which serves as the foundation for the generative AI responses in Yandex Search, is now available for researchers and developers to study and integrate into their own projects.
The model features a hybrid architecture combining Encoder-Decoder and Mixture of Experts (MoE) technologies. While the model contains 35 billion parameters, it activates less than 2% of them—approximately 600 million—per token processed. This design allows for high performance while minimizing computational requirements.
According to Yandex, the model outperforms compact alternatives like Qwen 3.5 2B/4B and T5 Gemma 2 4B-4B in Russian language tasks. It also demonstrates quality comparable to the Qwen 3.5 35B-A3B model, despite requiring significantly fewer resources for operation.
A fine-tuned version of this technology, known as Alice AI Search, has been integrated into Yandex Search since July. It provides concise answers directly below the search bar, trained on real user interaction data. Currently, the service reaches over 49 million users monthly.
