Sber Launches Latest AI Assistant, GigaChat Powered by Flagship Model

Sber has introduced a major update to its artificial intelligence (AI) assistant, GigaChat. This AI is now supported by the latest flagship model, GigaChat Ultra.

The latest version enables the AI assistant to remember facts about users to provide more personalised communication and solutions, independently search for information on the internet, and generate text responses twice as fast.

The launch of this new model opens opportunities not only for end-users but also for developers to create applied AI products and services using GigaChat Ultra.

Users can run code directly in the interface and obtain answers to questions about its own capabilities, based on the latest documentation.

“We are transforming from a mere tool that provides answers into a multi-agent AI assistant. Moreover, our goal is much greater: we are creating an era where conventional mobile applications will be replaced by neural network-based interfaces,” explained Sberbank’s Senior Vice President and Head of Generative AI Development, Anton Frolov, in a written statement on Friday (27/3/2026).

“Needed features will emerge on demand, making navigation in the digital world smoother. GigaChat Ultra is one of the largest models in the world that is fully developed and trained in Russia,” he added.

Frolov stated that the model can remember user preferences, work faster, understand tasks more deeply, and provide higher-quality recommendations. Sber has removed the last barriers in human-machine interaction.

Long-Term Memory

One of its main innovations is long-term memory. While contextual memory (short-term) is limited to a single conversation session and disappears after the session ends, GigaChat’s long-term memory works differently—this memory stores specific user facts across sessions and uses them in subsequent conversations.

GigaChat is capable of remembering the following:

● Hobbies, tastes, and interests;

● Profession, education, life goals, and habits;

● Personal data, only to the extent shared by the user themselves;

● Information about family members and pets.

The system automatically recognises important facts without burdening the memory with trivial matters, such as short-term plans or widely known general knowledge. All data is stored in a unified profile that is synchronised between the web version, mobile app, and Telegram bot via Sber ID login.

Users have full control over this feature: the memory can be enabled or disabled at any time in the settings.

Response Speed Doubled

GigaChat can now provide text responses twice as fast compared to Sber’s previous flagship model. This directly affects how quickly users receive answers, even for complex questions requiring in-depth analysis—the results appear almost instantly.

This speed improvement is achieved thanks to the Mixture of Experts (MoE) architecture. The model works like a team of specialist experts, where each expert handles a specific type of task.

Only the relevant ‘experts’ respond to a question, so the entire model does not work simultaneously.

Real-Time Conversation Mode

GigaChat can now automatically connect to internet searches to obtain up-to-date information, so users do not need to activate this option manually.

This ensures accurate responses when discussing current news, stock prices, and ever-changing data. The search function is also equipped with a special rephraser system, which reformulates the user’s question to improve relevance and the quality of the final answer.

Online search is now also available in voice communication mode. Conversations become fully interactive: users can interrupt the model, clarify details, or switch topics instantly—without delays in processing context changes.

After the conversation session ends, a complete transcript of the conversation is automatically saved.

GigaChat Has Self-Awareness Features

GigaChat is equipped with a self-awareness mechanism, allowing the model to accurately answer questions about its own characteristics. When responding to such questions, the model refers to the latest documentation containing the current version, available functionality, limitations, and typical behaviour.

In this way, common issues that often occur in language models, such as providing incorrect or outdated information about its capabilities—for example, claiming non-existent features or not recognising existing ones—can be avoided.

Code Interpreter: GigaChat as an Analytical Environment

With the integrated code interpreter, GigaChat functions as an isolated execution environment, allowing users to run software code directly within the assistant’s interface.

Before this function was introduced, the model could only write code and display it to the user; execution and testing of results had to be done with external tools.

Now, GigaChat can generate code and run it immediately in a secure sandbox, without affecting the user’s system.

The code interpreter also supports file uploads, performs advanced numerical calculations, validates data structures, and creates graphs and charts directly in the chat.

Thus, GigaChat becomes a complete analytical tool ideal for reports, tables, and processing large datasets.

Training Process

The training process took place in three stages. In the first stage, knowledge coverage was expanded by adding academic books, materials related to mathematics and programming, and increasing the volume of multilingual data—which now includes ten languages.

In the second stage, the focus was on improving specific skills: expanding the code corpus, with additional data including physi