Weibo's VibeThinker-1.5B AI Model Outperforms Larger Rivals

3 hours ago7 min read

In a development that challenges fundamental assumptions about artificial intelligence scaling, Weibo's AI division has released VibeThinker-1. 5B, a compact 1.5 billion parameter model that demonstrates remarkable reasoning capabilities despite its modest size. Built as a fine-tuned variant of Alibaba's Qwen2.5-Math-1. 5B and available under a permissive MIT license on Hugging Face, GitHub, and ModelScope, this model achieves what many considered impossible: outperforming giants like DeepSeek's 671-billion parameter R1 model on formal reasoning benchmarks while requiring only $7,800 in post-training compute resources.The secret lies not in brute-force parameter scaling but in Weibo's innovative Spectrum-to-Signal Principle training framework, which decouples supervised fine-tuning and reinforcement learning into distinct phases. During the 'Spectrum Phase,' the model learns to generate diverse solution paths rather than optimizing for single-answer correctness, while the subsequent 'Signal Phase' uses MaxEnt-Guided Policy Optimization to identify and amplify the most correct paths from this expanded solution space.This methodological breakthrough represents a significant departure from the industry's prevailing wisdom that larger models inevitably yield better performance, suggesting instead that sophisticated training techniques can enable smaller models to excel in specific domains. On benchmarks including AIME25 (74.4), LiveCodeBench v6 (51. 1), and GPQA-Diamond (46.7), VibeThinker-1. 5B competes with or surpasses models hundreds of times larger, though it understandably lags in general knowledge tasks where parameter count still matters.For enterprise applications, the implications are profound—this model can run efficiently on edge devices and mobile platforms while reducing inference costs by 20-70 times compared to larger alternatives. Weibo's strategic pivot into AI research reflects both the competitive pressures facing China's social media landscape and the company's ambition to leverage its substantial user data and technical resources beyond traditional social networking.As the AI field matures beyond simple scaling laws, VibeThinker-1. 5B demonstrates how targeted architectural innovations and training methodologies can create highly capable models that democratize access to advanced reasoning systems while dramatically reducing computational and financial barriers.

#Weibo

#VibeThinker-1.5B

#AI model

#open-source

#reasoning performance

#cost-effective

#featured

Stay Informed. Act Smarter.

Get weekly highlights, major headlines, and expert insights — then put your knowledge to work in our live prediction markets.

Comments

Loading comments...