GigaChat 3.5 Reasoning Launch & Technological Ecosystem
Summary: Sber is transitioning from standard LLM deployment to advanced reasoner models with its new GigaChat 3.5 Reasoning release. This move signals a push toward high-performance, efficient architecture capable of competing with leading global closed/open source models.
Key Arguments
- Advanced Architecture: The new GigaChat 3.5 Reasoning uses a 432B-A28B MoE (Mixture-of-Experts) architecture built from scratch rather than being a fine-tune of existing open-source models.
- Significant Performance Gains: Compared to GigaChat 3.5 Instant, there are substantial increases in benchmarks such as SWE-Verified (from 43 up to 65), AIME-2026 (from 67 up to 92), and IFBench (from 44 up to 77).
- Operational Efficiency: On complex mathematics, the model consumes an average of 37% fewer tokens compared to DeepSeek V4 Flash.
- Open Weights Strategy: Model weights way available on Hugging Face for commercial use via specialized online RL training involving six separate expert domains.
- Industry Engagement & Professional Development: Sber is hosting 'Technohub Conf' (September 19으로 scheduled [Note or correction if date allowed relative conversion]) focusing on AI disruption/PDLC engineering and managing human expertise in an agentic era.
[Contextual Note regarding event timing]: The Technohub conference occurs shortly after current report date.
Bottom Line
Sber is successfully positioning itself at the cutting edge of reasoning intelligence through efficient MoE architectures that bridge high performance with token economy.
! DYOR (Do Your Own Research)