In detail
- VibeThinker-3B matches DeepSeek V3.2 and Kimi K2.5 on benchmarks like AIME26, despite those models having 200–333× more parameters.
- On LeetCode contests (April–May 2026), the model solved 123 of 128 problems on first attempt—outperforming GPT-5.2, Qwen3-Max, and Kimi K2.5.
- Weakness in factual knowledge: on GPQA-Diamond benchmark, VibeThinker-3B falls well behind larger competitors.
- Built on Alibaba's Qwen2.5-Coder-3B; Sina's innovation is in post-training with specialized reinforcement learning for multi-step reasoning tasks.
Why it matters
The model demonstrates that reasoning capabilities compress efficiently, while factual knowledge requires scale. This has implications for cost optimization of AI systems in practice.
For you Watch whether smaller, specialized models become more cost-effective for your coding or math tasks—VibeThinker suggests you may not always need the largest models.