VibeThinker-3B, scaled from a 1.5B model, reaches frontier-level performance in math and coding tasks. It scores 94.3 on AIME'26, 80.2 on LiveCodeBench v6, 76.4 on IMO-AnswerBench, and 93.4 on IFEval, with 96.1% success on first-attempt LeetCode problems.
VibeThinker-3B achieves frontier math and coding performance
from English