Sber has released GigaChat 3.5 Reasoning, the first Russian open-source AI model capable of step-by-step reasoning. The model utilizes online reinforcement learning technology, allowing it to process complex task chains and self-correct during the generation process.
The reasoning mode is designed for tasks requiring multi-step logic, such as contract analysis, financial calculations, software debugging, and complex travel planning. Users can activate this feature manually via a dedicated button in the interface.
The new version shows significant performance improvements in benchmarks. In IFBench, scores increased from 44 to 77 points, in Natural Plan from 64 to 80, and in Live Code Bench v6 from 56 to 85. The model also consumes 37% fewer tokens for math tasks compared to DeepSeek V4 Flash Preview.
GigaChat 3.5 Reasoning is available to all users, with developers gaining access under the MIT license for product integration. A business-oriented API version is scheduled for release in the near future.
