VibeThinker-3B, a 3 billion parameter model, has been introduced as a compact dense model for enhanced verifiable reasoning.

The model leverages a Spectrum-to-Signal post-training paradigm with curriculum-based supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation. It scores 94.3 on AIME26, improving to 97.1 with test-time scaling, and 80.2 Pass@1 on LiveCodeBench v6. Additionally, it achieves a 96.1% acceptance rate on unseen LeetCode contests and scores 93.4 on IFEval. These results suggest that compact models can match or exceed larger flagship models like DeepSeek V3.2 and GLM-5.

Source: arxiv.org