Sber has introduced GigaChat 3.5 Reasoning, a multimodal AI model capable of breaking down complex tasks into sequential steps. The model autonomously plans solutions, utilizes external tools, and verifies its own intermediate results before providing a final answer.
The model is based on GigaChat 3.5 Ultra and was trained using reinforcement learning on mathematical and coding problems. It features a linear attention architecture designed to process long contexts more efficiently by retaining key information rather than re-evaluating the entire input.
In benchmark testing, the model showed significant performance gains. Scores in IFBench rose from 44 to 77, Natural Plan from 64 to 80, and Live Code Bench v6 from 56 to 85. Additionally, it consumes 37% fewer tokens on math tasks compared to DeepSeek V4 Flash Preview.
Users can activate the reasoning mode via a dedicated button. The model is currently available to all GigaChat users, while developers can access it for free under an MIT license. A business API version is scheduled for release in the near future.
