Sber has unveiled GigaChat 3.5 Reasoning, a new multimodal flagship model designed for complex, multi-step tasks. The model is currently available to GigaChat users, and developers can access it under an MIT open-source license. API access for business users is expected to launch in the near future.
The core feature of the model is its ability to break down tasks into sequential steps before generating an answer. It can create plans, identify missing information, utilize search tools, and verify intermediate results. If an error is detected, the model can adjust its strategy accordingly.
The reasoning mode is activated via a dedicated button. While simple queries are handled directly, complex tasks—such as coding, financial analysis, or logistics planning—trigger a deeper analysis process. The model was built upon GigaChat 3.5 Ultra, with training focused on multi-step problem-solving and automated verification of reasoning paths.
Performance metrics show significant improvements compared to the version without the reasoning mode. In the IFBench test, the score rose from 44 to 77 points, in Natural Plan from 64 to 80, and in Live Code Bench v6 from 56 to 85. Sber claims the model is competitive with leading international alternatives in instruction following and planning.
Sber reports that GigaChat 3.5 Reasoning is more token-efficient than some open-source competitors. Specifically, when solving mathematical problems, the model demonstrates a 37% average advantage in token usage compared to the DeepSeek V4 Flash Preview.
