🤖 Sber releases GigaChat 3.5 Reasoning — a reasoning model trained from scratch
On September 10, Sber published on Hugging Face the FP8 weights of a 432B-A28B MoE (28B active parameters) under the MIT license with commercial use rights. It is the first open reasoning model in Russia trained from scratch, not fine-tuned: post-training used online RL with six domain experts, combined into a single model through on-policy distillation. SWE-bench Verified increased from 42.6 to 64.7, and AIME-2026 from 67 to 92.
🌍 The open weights under MIT give companies sovereign reasoning inference without a third-party API, and the ~37% savings in reasoning tokens on math reduces the cost of complex tasks.
👤 The weights are open — run via SGLang or the reasoning mode in GigaChat. The RL pipeline is publicly detailed: six experts with different rewards.
Source 1: https://huggingface.co/ai-sage/GigaChat3.5-432B-A28B-Reasoning Source 2: https://habr.com/ru/companies/sberbank/articles/1080596/