VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO (arxiv.org)

VibeThinker-3B, a newly developed language model with 3 billion parameters, is making waves in the AI community by achieving top-tier reasoning capabilities typically reserved for much larger models. Leveraging a refined approach that incorporates curriculum-based fine-tuning, multi-domain reinforcement learning, and offline self-distillation, this compact model excels in demanding verifiable tasks. It scored 94.3 on the AIME26 benchmark and demonstrated remarkable generalization abilities with a 96.1% acceptance rate on unseen LeetCode contests, placing it on par with heavyweight models like DeepSeek V3.2 and Gemini 3 Pro. The significance of VibeThinker-3B lies in its demonstration that small models can achieve high performance without sacrificing instruction controllability. This challenges the prevailing notion that larger models are inherently superior by introducing the Parametric Compression-Coverage Hypothesis, which suggests that compact models may serve as effective alternatives or complements to larger systems in delivering robust reasoning capabilities. The findings could reshape the development and deployment strategies within the field, pushing researchers to explore the untapped potential of smaller models in offering sophisticated AI solutions.