DeepSeek’s MathV2 Just Surpassed Gemini 3 DeepThink

DeepSeek just dropped a new math model that pushes structured reasoning past Gemini 3 DeepThink, hitting Olympiad-tier proof accuracy with a full student–teacher–supervisor loop that checks and corrects its own logic.

At the same time, Tencent released HunyuanOCR, a tiny 1 B-parameter expert model that reads documents, receipts, tables, and multilingual pages better than massive vision-language systems built with tens of billions of parameters. ]


These two drops show how fast specialist models are overtaking giant general models in the tasks that actually matter.

AI Revolution
ai news