Weibo open-sources VibeThinker-3B; 3B model enters frontier reasoning rankings

ME News reports that on June 18 (UTC+8), according to monitoring by Beating, the Weibo team recently open-sourced the 3-billion-parameter inference model VibeThinker-3B. The model achieves state-of-the-art performance on tasks such as mathematics and programming, with some metrics approaching or surpassing large flagship models like DeepSeek V3.2, GLM-5, and Gemini 3 Pro. VibeThinker-3B is built upon Qwen2.5-Coder-3B and undergoes secondary training using the Spectrum-to-Signal process: the model first accumulates problem-solving patterns by progressing from easy to difficult problems, then amplifies signals of correct solutions through reinforcement learning, utilizing a 64K-step reasoning space throughout training to prevent interruption of reasoning steps. For mathematics and programming, the model一方面 collects its own high-quality correct steps for self-distillation and imitation, and另一方面 introduces step-level evaluation during problem-solving for self-verification. This self-verification mechanism increased the model’s score on the AIME26 math test from 94.3 to 97.1. The development team proposes the “Parameter Compression-Coverage Hypothesis” in their report, suggesting that logical reasoning is a highly compressible capability, primarily reliant on rules and error correction, enabling top-tier performance with only a 3B-parameter model; whereas open-domain knowledge requires massive parameters for rote memorization. Due to its parameter scale, VibeThinker-3B still lags behind larger models in covering common-sense open-domain knowledge. The team emphasizes that the goal is not to replace large models with small ones, but to explore the capability boundaries of compact models under well-defined verification mechanisms. (Source: BlockBeats)