Sber launches GigaChat 3.5 Ultra with faster long-text handling and stronger coding

Sber has released GigaChat 3.5 Ultra, a new flagship AI model now free in the GigaChat assistant and open source for developers. The model is designed to improve coding, long-document analysis and agent-style workflows while using fewer computing resources than the prior version.

Why it matters: - GigaChat 3.5 Ultra is aimed at practical AI work, not just chat. That puts coding, document analysis and automated multi-step tasks in reach for more users and developers. - Sber says the model is nearly half as compact as the previous version, which could make independent deployment cheaper and easier on lower-cost hardware. - The release extends access in two directions at once: free consumer use inside the GigaChat assistant and open-source availability for builders.

What happened: - Sber introduced GigaChat 3.5 Ultra on July 7, 2026. - The model is available free of charge in the GigaChat AI assistant. - The model is also open source for developers worldwide. - Sber positions GigaChat 3.5 Ultra as its new flagship model.

The details: - GigaChat 3.5 Ultra is built to generate long text up to four times faster than before. - The model uses fewer resources than the prior version. - Sber says the model is nearly half as compact as the previous GigaChat Ultra. - The model performs better on coding, mathematics, long-text work and autonomous agent scenarios. - GigaChat 3.5 Ultra can generate and verify code with more confidence. - The model solves mathematical problems and financial calculations with higher accuracy. - The model produces responses that are more precise, better structured and easier to understand. - GigaChat 3.5 Ultra can analyze contracts, technical regulations, reports and other long documents without losing context. - The model uses linear attention, which accumulates context instead of reprocessing the full text each time. - Sber says the model can let a user assign a task, then independently find information, write and execute code, access a service and return a finished result. - That workflow could automate monitoring, data processing and scheduled report generation. - In tests on programming tasks, math problems, complex multi-step assignments and Russian-language dialogue, GigaChat 3.5 Ultra outperformed Sber’s previous flagship model. - Sber says the model also came close on several metrics to strong open models such as DeepSeek 3.2 while staying nearly half as compact. - Anton Frolov, senior vice president and head of GenAI Development at Sberbank, said the company has more than doubled its experiments to 1,500. - Sber says the model is based on a proprietary domestic architecture with linear attention technology developed by Sber’s team. - The model is one of the largest linear-attention models released in open source. - Training focused on natural, human-generated texts that were multi-level classified and filtered. - The model uses a Mixture of Experts architecture. - The smaller size is intended to reduce compute needs and make the model easier to run on affordable hardware.

Between the lines: - The release signals Sber’s push to compete on efficiency as much as model quality. - By emphasizing compactness, linear attention and open source access, Sber is making a case that domestic infrastructure can support serious AI development with fewer resources. - The focus on agent-like behavior suggests Sber wants GigaChat to move deeper into enterprise automation and developer workflows.

What's next: - Developers can now build services and AI agents on top of the open-source model. - More companies may be able to run GigaChat 3.5 Ultra independently because of its smaller footprint and lower hardware requirements. - Sber appears to be positioning the model as a base for future products and research beyond the company. - More details are available on the company's websites and Sberbank.