Sber has unveiled a new AI model , GigaChat 3.5 Ultra , which, according to the company, delivers superior performance in programming tasks, mathematical calculations, and the analysis of large volumes of text. The model is available free of charge to both GigaChat users and developers as an open-source solution.
The company claims that GigaChat 3.5 Ultra is capable of generating long texts up to four times faster than the previous version while using fewer computational resources. The new model is also nearly half the size of its predecessor.
The model is based on an architecture developed by Sber that utilizes Linear Attention technology. This enables more efficient processing of long documents by gradually accumulating context, rather than re-analyzing the entire text from scratch.
Key improvements
Programming and Mathematics
According to Sber, the model generates and verifies code more accurately, solves mathematical problems and performs financial calculations more efficiently, and handles numerical data better. The company notes that the technology can be useful to developers, analysts, and engineers in their day-to-day work.
Analysis of large documents
The new model is designed to analyze contracts, technical regulations, reports, and other voluminous documents. Sber claims that, thanks to a linear attention architecture, document processing speed has increased up to fourfold while maintaining accuracy in context comprehension.
AI agent support
GigaChat 3.5 Ultra is also designed for developing autonomous AI agents. The model is capable of independently searching for information, generating and executing code, interacting with external services, and preparing final results, thereby automating various routine processes such as data processing, monitoring, and the preparation of regular reports.
Comparison with other models
According to Sber, internal tests show that GigaChat 3.5 Ultra outperforms the company's previous flagship model in tasks involving programming, mathematics, multi-step logical reasoning, and Russian-language dialogue.
The company also notes that, across several performance metrics, the new model has come close to one of the most popular open-source large language models— DeepSeek 3.2 —while maintaining an architecture nearly half the size.
Source: giga.chat
0 Comments