July 7 : Russians now have access to the new flagship model, GigaChat 3.5 Ultra. It has become smarter, generates long text up to four times faster, consumes fewer resources, and is nearly half as compact as the previous version. The updated model handles tasks related to coding, mathematics, working with long texts, and autonomous agent scenarios more effectively. The model is based on a proprietary domestic architecture featuring linear attention technology, developed by Sber’s team.
The improved model is available to anyone. In the GigaChat AI assistant, it is accessible to all users who want to apply AI for personal and work‑related tasks. In open source, it is freely available to developers worldwide for integration into their services and for building AI agents.
What GigaChat 3.5 Ultra can do?
Coding and mathematics – the model generates and verifies code with greater confidence, solves mathematical problems and financial calculations with higher accuracy, and handles numbers more reliably. Responses have become more precise, better structured, and easier to understand. All of this enables the model to be integrated into the real‑world workflows of developers, analysts, and engineers.
Reading and analyzing long texts – it efficiently analyzes contracts, technical regulations, reports, and other lengthy documents without losing accuracy or context. Thanks to the linear attention architecture, it does not re‑read the text from the beginning each time; instead, it gradually accumulates context, much like a person who remembers the gist of a conversation. The speed of working with long texts has increased up to fourfold.
Autonomy and AI agents – you can assign a task to the model, and it will independently find information, write and execute code, access the required service, and return a ready‑made result. This makes it possible to automate routine tasks: monitoring, data processing, and scheduled report generation.
In tests measuring the AI’s ability to solve programming tasks, mathematical problems, complex multi‑step assignments, and the quality of Russian‑language dialog, GigaChat 3.5 Ultra outperformed Sber’s previous flagship model. And in several metrics, it came close to the results of strong open models, such as DeepSeek 3.2, while being nearly half as compact.
Anton Frolov, senior vice president, head of GenAI Development, Sberbank:
“We are living in a time when the gap between human capabilities and AI potential is shrinking rapidly. GigaChat 3.5 Ultra is our step toward what an AI tool for real‑world tasks should be: a full‑fledged partner capable of thinking within the logic of a specific process, not just answering questions. To develop such a model, you need to constantly experiment and try things no one has done before – the number of our experiments has more than doubled, reaching 1,500. We have proven that it is possible to build a strong model using a proprietary architecture and with fundamentally fewer resources. We want our solutions to become the foundation for new products and research that go far beyond Sber.”
GigaChat 3.5 Ultra is fully Sber’s brainchild: the team created a unique architecture leveraging linear attention technology. Unlike the classic attention mechanism in AI models, which re‑checks each new word against the entire preceding text every time, linear attention memorizes the essence of what has been read once and then simply adds to that memory – roughly like a person who keeps a brief summary of a book in mind rather than flipping back to the first page with every new page. GigaChat 3.5 Ultra is one of the largest models with linear attention among those released in open source.
During training, the focus was on natural, human‑generated texts that underwent multi‑level classification and filtering. The expanded dataset collected enabled the achievement of better metrics. The model follows a MoE architecture and is approximately half the size of the previous GigaChat Ultra version, which reduces computational resource consumption and allows it to be deployed on more affordable hardware – meaning more companies and developers will be able to run the model independently.