GigaChat 3.5 Reasoning is a new open-source LLM designed to reason before generating responses. The model breaks problems into stages, builds execution plans, checks intermediate results, and self-corrects when needed.

Built on GigaChat 3.5 Ultra, it was trained on math and coding tasks using multiple step-by-step reasoning paths. An automated verification step reinforces the paths that lead to correct answers, enabling the model to plan multi-step actions, decide when to call external tools, and revise earlier steps independently.

The model uses a proprietary linear attention architecture, which improves efficiency on long contexts by retaining key processed points rather than re-matching queries against the entire prior text.

On math problems, GigaChat 3.5 Reasoning uses on average 37% fewer tokens than DeepSeek V4 Flash Preview. Benchmark gains over the non-reasoning version:
• IFBench: 44 → 77
• Natural Plan: 64 → 80
• LiveCodeBench v6: 56 → 85

The model is open-sourced under the MIT license. Weights are available on Hugging Face:  fp8 | bf16