Sberbank's AI Sage has released GigaChat-3.5 Reasoning, a 432B-A28B Mixture of Experts (MoE) model utilizing Gated DeltaNet for long-context efficiency.
The model was trained using domain experts in code, math, and general tasks via CISPO, then distilled into a single model through on-policy distillation. In evaluations, it performs close to DeepSeek V4 Flash Preview while using 37% fewer tokens in its reasoning traces.
Weights are available under the MIT license on Hugging Face, and the model can be tested at giga.chat.