Yandex: Alice AI Family of generative models

The main articles are:

• Speech Recognition (Technology, Market)
• Speech technology: On the path from recognition to understanding
• LLM (Large Language Models)

Alice AI is a family of generative models that includes optimized Yandex models - Alice AI LLM, Alice AI Art and Alice AI VLM.

2025

At the heart of Alice's AI chat

in and Alice messengers Telegram MAX now runs on Alice AI. Chat boat Alice AI gives more accurate and complete answers, providing them with links to illustrations. Yandex announced this on December 11, 2025. The user can solve his problem right in the messenger and send the answer to neuronets someone from contacts or to a group. You can send a request to a chatbot in writing or by voice in a message. More here.

At the heart of the Yandex.Maps AI chat

Yandex Maps has a chat with AI based on Alice AI. It is available to all users - just click on the "Ask AI" icon under the active search bar. Using AI, you can communicate with Maps: ask what to do in the city, receive collections and reviews of various places, or even ask to plan a day on a trip. Yandex announced this on December 11, 2025. Read more Yandex.Maps#.2A_.D0.97.D0.B0.D0.BF.D1.83.D1.81.D0.BA_.D0.98.D0.98-.D1.87.D0.B0.D1.82.D0.B0 here.

Open business access to the Alice AI family of generative models

Yandex B2B Tech opens up business access to the Alice AI family of generative models. The Alice AI neural network is already working on them - now the business will be able to use them for its tasks and create its own AI applications on their basis. Yandex B2B Tech announced this on November 25, 2025. Compared to the previous generation of neural networks, the Alice AI LLM model is more universal for business: it has become smarter and works better in dialogue mode. She will be able to explain complex things in simple language, given the specifics of the business, internal documents and instructions. Alice AI LLM can already be used on the Yandex AI Studio platform, Alice AI ART will also be available on it soon.

According to Yandex B2B Tech, Alice AI LLM in 60% of cases surpasses neuronets DeepSeek V3.1 and - Qwen3 235b in the quality of solving business problems. In comparison with them, it performs better when generating answers by documents, writing texts and editing them in a different key or different styles, as well as in answers to general questions.

Alice AI LLM is Yandex's flagship model that has undergone a full cycle of training on the company's own data, from loading general world data (pretrane) to adapting the model to specific tasks (SFT). Even before all these stages, the weights of open source models were used as the initial parameters of the model - this made it possible to reduce the time for developing the model. At the same time, the basis of the final model is formed already at the next, more difficult stages of training. As a result, all original model parameters were recalculated and changed based on Yandex data.

Alice AI LLM has retained all the strengths of YandexGPT: more than 40 thousand of our customers use it for concise and clear answers, accurate extraction of data from documents, search for information in RAG scenarios. At the same time, we strengthened the model in those tasks where the business needs a detailed and "human" answer: for example, explain the specifics of a complex product to customers, draw up answers to possible customer objections for the sales department, or prepare educational materials on corporate knowledge bases. This adds many new scenarios for using the model, which are in demand among our customers, "explained Artur Samigullin, head of Yandex AI Studio at Yandex B2B Tech.

As part of the beta testing, dozens of companies gained access to Alice AI. Just AI has already tested the model in agency systems and RAG scenarios, where AI can perform complex tasks: search for data, analyze context, build action chains. The company presentsimple.ai used the model to generate presentations - they noted that Alice AI demonstrates higher creativity, for example, it offers non-standard options for slide titles.

At the same time, the cost of using Alice AI can be significantly lower than that of open source neural networks, thanks to the separation of text into tokens optimized for Russian. A single token in Alice AI places about 4-5 characters in Cyrillic, and in open source models about 2-3 characters. Thus, the final cost of using Alice AI will be 1.5-2 times lower than open source models with the same tariff. In addition, now user requests to the neural network and model responses are charged differently - tokens "at the entrance" of the neural network are four times cheaper than "at the exit," which allows you to use the model more economically.