Sina's open model VibeThinker-3B aims to show reasoning compresses well but factual knowledge doesn't

Sina Weibo has launched the VibeThinker-3B model, featuring three billion parameters, which has outperformed models like DeepSeek V3.2 and Kimi K2.5, despite those models being up to 333 times larger in size. The effectiveness of VibeThinker-3B is attributed to its multi-stage training technique, demonstrating that logical reasoning can be compressed efficiently in smaller models, while extensive general knowledge does not compress as effectively.