GigaChat 3.5 Reasoning is a newly developed open-source large language model (LLM) designed to enhance its response process by incorporating a thoughtful approach. This innovative model breaks down challenges into manageable stages, formulates a strategy, verifies interim results, and makes adjustments as necessary. Built on the foundation of GigaChat 3.5 Ultra, it navigates through various reasoning pathways for tasks in mathematics and programming, employing automated verification to strengthen the accuracy of its answers.
One of its standout features is a proprietary linear attention mechanism that optimizes performance with lengthy contexts, enabling the model to retain essential information without needing to reassess from the beginning. Additionally, it demonstrates token efficiency, utilizing 37% fewer tokens than the DeepSeek V4 Flash Preview when tackling mathematical problems.
When compared to its non-reasoning counterpart, GigaChat 3.5 shows significant benchmark improvements across various tests: • IFBench: increased from 44 to 77 • Natural Plan: improved from 64 to 80 • LiveCodeBench v6: surged from 56 to 85
The model is available under the MIT license, with weights accessible on Hugging Face in both fp8 and bf16 formats.
Informational material. 18+.
