Sber Presents GigaChat 3.5 Reasoning with Reasoning Mode

Sberbank has released its flagship language model, GigaChat 3.5 Reasoning, equipped with a specialized sequential reasoning mode. The new version can break down complex tasks into stages, formulate a solution plan, and independently verify intermediate results to correct errors.

Efficient Resource Utilization

A key feature of the release is the optimization of token consumption. When performing mathematical calculations, the new model uses, on average, 37% fewer tokens compared to the open-source AI model DeepSeek V4 Flash Preview.

Benchmark Test Results

Enabling the reasoning mode significantly improved the model's performance across various tests. The results demonstrate a substantial increase in efficiency compared to the version without this functionality.

• IFBench: score increased from 44 to 77 points
• Natural Plan: rating rose from 64 to 80 points
• Live Code Bench v6: result jumped from 56 to 85 points

Availability and Integration

The model is available free of charge under the MIT license. Developers can integrate it into their own products by downloading files from HuggingFace and GitVerse platforms. The updated mode is also available for testing on the official giga.chat website.