In short. Alice AI Search generates answers in Yandex Search using a fine-tuned model whose pretrained base is released separately.
Definition: Alice AI Search is Yandex's language model for producing AI answers within Yandex Search. Yandex's September 2026 announcement describes it as a fine-tuned version of the base model released as Alice AI Search Pretrain.
Key facts
- Yandex's model announcement reports that Alice AI Search runs in Search from July 2026 and is trained further using search interactions and real search examples.
- As of October 2026, Yandex's Search help describes answers above search results, with source links and conversational follow-up questions.
- Yandex's Webmaster documentation says these answers use pages indexed by Yandex and documents a robots.txt restriction for excluding page content.
What does the model do?
Yandex's September announcement assigns Alice AI Search the task of producing concise, easy-to-follow search answers. It places those answers directly below the search bar.
The announcement separates the deployed, fine-tuned model from Alice AI Search Pretrain, the released base model. The latter is available on Hugging Face under the Apache 2.0 license for study and further development.
Its encoder-decoder architecture separates processing source material from generating an answer. Yandex describes the model as designed to handle large request volumes with relatively low computing requirements.
How do search answers use web pages?
Yandex's Search help describes a process in which search algorithms select relevant web materials, a neural network analyzes them in real time, and the system compiles an answer with links to those materials.
For complex questions, the feature breaks the request into parts before combining information. The documentation gives shade-tolerant plants that need no daily watering as an example. Follow-up questions preserve the previous question's context.
The same help page distinguishes school math problem-solving mode: it presents solution steps from memory and omits source links. This exception matters when checking whether an answer refers to a website.
What controls does a website owner have?
Yandex's Webmaster guidance identifies indexed pages as the material used for answers. It recommends well-structured, informative and grammatically correct content.
A robots.txt file contains crawler access directives. Yandex documents a restriction for YandexAdditional or YandexAdditionalBot to exclude page content from these answers. The following lines reproduce the operative part of its example:
User-agent: YandexAdditionalBot
Disallow: /cats/mycat.html
The example targets a particular page path. The documentation says the change takes effect after Yandex robots recrawl the page.
Related terms
FAQ
Is Alice AI Search Pretrain the deployed search model?
Yandex's announcement identifies Pretrain as the base model and Alice AI Search as its fine-tuned search version. The public release provides a foundation for further development.
How can you open the pages used in an answer?
Yandex's Search help instructs you to select a source link from the answer's list. Its school math mode uses a different process without those links.
Sources
- Yandex, "Yandex Open-Sources the Language Model Behind Its AI Answers in Search", 14 September 2026
- Yandex Search Help, "Search with Yandex AI", undated
- Yandex Webmaster Help, "Search with Yandex AI", undated
Check how a page reads to answer engines. Open the GEO score analyzer