Yandex Open-Sources the Language Model Behind Its AI Answers in Search

Yandex has open-sourced Alice AI Search Pretrain, the base language model behind the system used to generate AI answers in Yandex Search. The model was trained from scratch and is designed to combine strong answer quality with efficient, high-throughput inference.

Alice AI Search Pretrain is available on Hugging Face under the Apache 2.0 license. Yandex is the first company to release the pretrained base version of a model used to generate AI answers in search.

The model combines an encoder-decoder architecture with Mixture of Experts (MoE), an approach that has until recently been used primarily in research. It has 35 billion parameters in total, but only around 600 million — less than 2% — are activated when processing each token.

This makes the model more efficient, allowing it to handle a large volume of requests with relatively low compute requirements. The encoder-decoder architecture separates source processing from answer generation, while MoE activates only the subset of model parameters needed for a given token.

Alice AI Search Pretrain has already completed pretraining and has broad general knowledge. Developers and researchers can study its architecture, experiment with the model and use it as a foundation for their own projects.

In Yandex's blind tests of Russian-language answer quality, Alice AI Search Pretrain outperformed Qwen3.5-2B, Qwen3.5-4B and T5 Gemma 2 4B-4B Base. Its results were also comparable to Qwen3.5-35B-A3B, a model that requires considerably more computing resources.

A fine-tuned version of the model, Alice AI Search, has been running in Yandex Search since July 2026 as part of the Alice AI family of generative models. Yandex fine-tuned it using signals from how people interact with Search and real examples of search use, with the aim of producing concise, easy-to-follow answers.

These AI answers appear directly below the search bar, helping users get the key information on a topic without having to open multiple results. More than 49 million people use the feature each month, making it Yandex's most widely used generative AI product.