VibeThinker-3B: 3B model beats Opus 4.5 in reasoning
A 3B-parameter model outperforms much larger systems on verifiable tasks
Researchers (Sen Xu et al.) introduced VibeThinker-3B, a 3-billion-parameter dense model that explores verifiable reasoning in small models. It uses an optimized pipeline with curriculum-based supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation. Scores: 94.3 on AIME26 (97.1 with claim-level test-time scaling), 80.2 Pass@1 on LiveCodeBench v6, 96.1% acceptance on recent unseen LeetCode contests, and 93.4 on IFEval. It matches or exceeds flagship models orders of magnitude larger like DeepSeek V3.2, GLM-5, and Gemini 3 Pro. The paper introduces the Parametric Compression-Coverage Hypothesis, suggesting verifiable reasoning is compressible into compact reasoning cores, while open-domain knowledge needs broad parameter coverage.
- VibeThinker-3B is a 3B-parameter model achieving frontier-level reasoning, matching systems 10x+ larger (e.g., DeepSeek V3.2, GLM-5).
- Scores 94.3 on AIME26 (97.1 with scaling), 80.2 Pass@1 on LiveCodeBench v6, and 96.1% on unseen LeetCode contests.
- Introduces the Parametric Compression-Coverage Hypothesis, redefining how compact models can achieve high-performance reasoning.
Why It Matters
Proves compact AI models can rival massive ones in reasoning, enabling efficient deployment without sacrificing performance.