GigaChat 3.5 Reasoning: benchmark results
Sber's separately trained GigaChat 3.5 reasoning checkpoint, a 432B mixture-of-experts model with 28B active parameters trained with online reinforcement learning and distillation. Provider: Sber. Released 2026-09-10. Access: Open.
Aggregate Rating 1721 ± 36, rank #159 of 1640 rated models, from 35 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| GigaChat Reasoning Card - AIME 2025, mean@32 | 89 | Score (%) | 100 |
| GigaChat Reasoning Card - AIME 2026, mean@32 | 92 | Score (%) | 100 |
| GigaChat Reasoning Card - Arena Hard Ru | 60.7 | Score (%) | 100 |
| GigaChat Reasoning Card - IFBench | 77 | Score (%) | 100 |
| GigaChat Reasoning Card - MERA-2.0 | 42.3 | Score (%) | 100 |
| GigaChat Reasoning Card - Ru LLM Arena | 64 | Score (%) | 100 |
| GigaChat Reasoning Card - StructEval | 85 | Score (%) | 100 |
| MERA v2 - SAGE | 68.8 | Score (%) | 89.7 |
| MERA v2 - RussianRegions | 14.5 | Score (%) | 87.2 |
| MERA v2 - Culture Specific | 41.2 | Score (%) | 84.6 |
| MERA v2 - RUBIN | 70.1 | Score (%) | 84.6 |
| MERA v2 - IFHardBench | 62.8 | Score (%) | 79.5 |
Interactive version: theaggregate.ai/model?slug=gigachat-3-5-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-16.