The A.I. Daily
GigaChat-3.5-Reasoning released
The team released GigaChat-3.5 Reasoning, a 432B-A28B MoE with Gated DeltaNet and on-policy distillation of domain experts into one model. They report it approaches DeepSeek V4 Flash Preview in evaluations while using 37% fewer tokens in reasoning traces; weights are on Hugging Face under an MIT license and a demo is available at giga.chat.