Sber has introduced a major update to its artificial intelligence\n(AI) assistant, GigaChat. This AI is now supported by the latest\nflagship model, GigaChat Ultra.<\/p>\n
The latest version enables the AI assistant to remember facts about\nusers to provide more personalised communication and solutions,\nindependently search for information on the internet, and generate text\nresponses twice as fast.<\/p>\n
The launch of this new model opens opportunities not only for\nend-users but also for developers to create applied AI products and\nservices using GigaChat Ultra.<\/p>\n
Users can run code directly in the interface and obtain answers to\nquestions about its own capabilities, based on the latest\ndocumentation.<\/p>\n
\u201cWe are transforming from a mere tool that provides answers into a\nmulti-agent AI assistant. Moreover, our goal is much greater: we are\ncreating an era where conventional mobile applications will be replaced\nby neural network-based interfaces,\u201d explained Sberbank\u2019s Senior Vice\nPresident and Head of Generative AI Development, Anton Frolov, in a\nwritten statement on Friday (27\/3\/2026).<\/p>\n
\u201cNeeded features will emerge on demand, making navigation in the\ndigital world smoother. GigaChat Ultra is one of the largest models in\nthe world that is fully developed and trained in Russia,\u201d he added.<\/p>\n
Frolov stated that the model can remember user preferences, work\nfaster, understand tasks more deeply, and provide higher-quality\nrecommendations. Sber has removed the last barriers in human-machine\ninteraction.<\/p>\n
Long-Term Memory<\/p>\n
One of its main innovations is long-term memory. While contextual\nmemory (short-term) is limited to a single conversation session and\ndisappears after the session ends, GigaChat\u2019s long-term memory works\ndifferently\u2014this memory stores specific user facts across sessions and\nuses them in subsequent conversations.<\/p>\n
GigaChat is capable of remembering the following:<\/p>\n
\u25cf Hobbies, tastes, and interests;<\/p>\n
\u25cf Profession, education, life goals, and habits;<\/p>\n
\u25cf Personal data, only to the extent shared by the user\nthemselves;<\/p>\n
\u25cf Information about family members and pets.<\/p>\n
The system automatically recognises important facts without burdening\nthe memory with trivial matters, such as short-term plans or widely\nknown general knowledge. All data is stored in a unified profile that is\nsynchronised between the web version, mobile app, and Telegram bot via\nSber ID login.<\/p>\n
Users have full control over this feature: the memory can be enabled\nor disabled at any time in the settings.<\/p>\n
Response Speed Doubled<\/p>\n
GigaChat can now provide text responses twice as fast compared to\nSber\u2019s previous flagship model. This directly affects how quickly users\nreceive answers, even for complex questions requiring in-depth\nanalysis\u2014the results appear almost instantly.<\/p>\n
This speed improvement is achieved thanks to the Mixture of Experts\n(MoE) architecture. The model works like a team of specialist experts,\nwhere each expert handles a specific type of task.<\/p>\n
Only the relevant \u2018experts\u2019 respond to a question, so the entire\nmodel does not work simultaneously.<\/p>\n
Real-Time Conversation Mode<\/p>\n
GigaChat can now automatically connect to internet searches to obtain\nup-to-date information, so users do not need to activate this option\nmanually.<\/p>\n
This ensures accurate responses when discussing current news, stock\nprices, and ever-changing data. The search function is also equipped\nwith a special rephraser system, which reformulates the user\u2019s question\nto improve relevance and the quality of the final answer.<\/p>\n
Online search is now also available in voice communication mode.\nConversations become fully interactive: users can interrupt the model,\nclarify details, or switch topics instantly\u2014without delays in processing\ncontext changes.<\/p>\n
After the conversation session ends, a complete transcript of the\nconversation is automatically saved.<\/p>\n
GigaChat Has Self-Awareness Features<\/p>\n
GigaChat is equipped with a self-awareness mechanism, allowing the\nmodel to accurately answer questions about its own characteristics. When\nresponding to such questions, the model refers to the latest\ndocumentation containing the current version, available functionality,\nlimitations, and typical behaviour.<\/p>\n
In this way, common issues that often occur in language models, such\nas providing incorrect or outdated information about its\ncapabilities\u2014for example, claiming non-existent features or not\nrecognising existing ones\u2014can be avoided.<\/p>\n
Code Interpreter: GigaChat as an Analytical Environment<\/p>\n
With the integrated code interpreter, GigaChat functions as an\nisolated execution environment, allowing users to run software code\ndirectly within the assistant\u2019s interface.<\/p>\n
Before this function was introduced, the model could only write code\nand display it to the user; execution and testing of results had to be\ndone with external tools.<\/p>\n
Now, GigaChat can generate code and run it immediately in a secure\nsandbox, without affecting the user\u2019s system.<\/p>\n
The code interpreter also supports file uploads, performs advanced\nnumerical calculations, validates data structures, and creates graphs\nand charts directly in the chat.<\/p>\n
Thus, GigaChat becomes a complete analytical tool ideal for reports,\ntables, and processing large datasets.<\/p>\n
Training Process<\/p>\n
The training process took place in three stages. In the first stage,\nknowledge coverage was expanded by adding academic books, materials\nrelated to mathematics and programming, and increasing the volume of\nmultilingual data\u2014which now includes ten languages.<\/p>\n
In the second stage, the focus was on improving specific skills:\nexpanding the code corpus, with additional data including physi<\/p>", "url": "https:\/\/jawawa.id\/newsitem\/sber-launches-latest-ai-assistant-gigachat-powered-by-flagship-model-1775217045", "image": "" }, "sponsor": "Okusi Associates", "sponsor_url": "https:\/\/okusiassociates.com" }