CHINESE SOCIAL MEDIA COMPANY POSTS A 3 BILLION PARAMETER MODEL THAT JUST OUTSCORED GPT AND CLAUDE AT MATH Nine researchers at Sina Weibo, the Chinese social platform most Americans have never thought about, posted a paper this week that is making the AI research community argue about benchmarks again. Their model, VibeThinker-3B, has 3 billion parameters, small enough to run on a consumer laptop, and it scored 94.3 on the 2026 American Invitational Mathematics Examination. On LeetCode coding contests it passed 123 out of 128 first-attempt submissions, a 96.1 percent rate that beat scores recorded by GPT-5.2, Claude Opus 4.6, and two Chinese flagship models under identical evaluation conditions. The catch, which researchers are pointing out loudly, is that benchmark performance and real-world usefulness are not the same thing. There is a growing and familiar gap between what these models claim to do on controlled tests and what users actually experience when they try to get work done. Still, the result forces an uncomfortable question for the major labs: if a team at a social media company in Beijing can match frontier reasoning performance at a fraction of the compute cost, what exactly is the moat that billions in training spend is supposed to protect? That question does not have a comfortable answer right now. Keywords: VibeThinker-3B, Sina Weibo AI model, small language models, AI benchmark 2026