GigaChat Ultra expands enterprise AI capabilities with memory, code execution and real-time search

Updated assistant integrates long-term memory, faster response generation and built-in analytical tools for business and developer use

Sber has released a new version of its AI assistant, GigaChat, now powered by the flagship model GigaChat Ultra, expanding its functionality for both end users and enterprise applications. The update focuses on improving how the system processes information, interacts with users and supports development environments, positioning the assistant as a broader interface for accessing and managing digital tasks.

One of the main additions is a long-term memory capability. Unlike session-based context, the system can retain user-provided information across interactions, including preferences, professional background and recurring patterns of use. This data is stored within a unified profile and can be managed or disabled by the user. According to the company, the goal is to enable more consistent interactions over time without relying on repeated inputs.

The model also introduces performance gains, with response generation occurring at approximately twice the speed of the previous version. This improvement is enabled by a Mixture of Experts architecture, where specialized components are activated depending on the task, rather than processing all queries through a single pathway.

GigaChat Ultra is designed to operate with real-time data access. The assistant can independently retrieve and process online information during interactions, allowing it to respond to queries involving dynamic content such as news updates or market data. This functionality is also available in voice mode, where users can interrupt, ???? adjust or redirect conversations without resetting context.

Another feature is an embedded code interpreter, which allows users to write, execute and validate code directly within the interface. The system can process uploaded files, perform calculations and generate visual outputs such as charts, turning the assistant into a self-contained analytical environment. This reduces reliance on external tools for testing and iteration.

The model also includes a self-referencing mechanism that enables it to provide accurate information about its own capabilities based on current documentation. This is intended to reduce inconsistencies in how AI systems describe their functions and limitations.

From a development perspective, Sber is making the model available for broader use, allowing organizations to deploy and adapt it within private environments. This includes the ability to integrate proprietary data and build custom applications on top of the model.

“We are taking a step from being just an answer-giving tool to becoming a multi-agent AI assistant,” said Anton Frolov, Senior Vice-President and Head of Generative AI Development at Sberbank. “Our goal is to move toward interfaces where functions are generated on demand, reducing the need for traditional applications.”

The training process for GigaChat Ultra included expanded multilingual datasets, technical and scientific materials, and domain-specific content such as finance and cybersecurity. According to the company, improvements were observed in reasoning tasks, structured responses and domain applicability.

With this release, Sber positions GigaChat Ultra as a platform not only for conversational use, but also for development, data analysis and system interaction, reflecting a broader shift toward AI-driven interfaces in enterprise environments.