Sou Hu Cai Jing·2025-11-13 21:18Core Insights - Weibo has launched its first open-source large model, VibeThinker-1.5B, claiming that smaller models can also exhibit high intelligence [1][2] - The model, with 1.5 billion parameters, challenges the notion that only models with massive parameter counts can achieve high performance, demonstrating that innovative algorithm design can yield significant results [2][5] Model Performance - VibeThinker-1.5B outperformed the DeepSeek-R1-0120 model, which has 671 billion parameters, in three challenging mathematics test sets (AIME24, AIME25, HMMT25) [2][5] - Its performance is comparable to the MiniMax-M1 model, which has 456 billion parameters, and it achieved similar results in the LiveCodeBench v6 programming test set against models with significantly higher parameters [2][5] Training Methodology - The model's success is attributed to the "spectrum to signal principle" (SSP) training method, which encourages exploration of all possible solutions during the learning phase, followed by reinforcement learning for efficient strategy optimization [5][6] - The post-training cost for VibeThinker-1.5B is less than $8,000, significantly lower than the costs for DeepSeek-R1 and MiniMax-M1, which are $290,000 and $530,000 respectively [6] Accessibility and Impact - The open-source nature of VibeThinker-1.5B aims to provide a cost-effective research path for medium-sized enterprises and academic research teams with limited computational resources, promoting inclusivity in cutting-edge model training [6]