VibeThinker-3B matches much larger models with 3B parameters

VibeThinker-3B, a 3.09B parameter model based on Qwen2.5-Coder-3B, reportedly competes with models 200-300x larger on math, coding, and reasoning. Built by Sina Weibo researchers using a novel post-training pipeline (SFT, MGPO, self-distillation), training cost estimated at $25k-$60k. Benchmarks are preliminary and need practical verification.