Weibo Open Sources VibeThinker-3B, Empowering the 3B Small Model to Advance into the Vanguard of Inference

According to DynoMinds monitoring, the Sina Weibo team recently open-sourced a 3 billion-parameter inference model named VibeThinker-3B.

The model has achieved state-of-the-art performance in mathematical and programming tasks, with some metrics approaching or surpassing flagship models such as DeepSeek V3.2, GLM-5, and Gemini 3 Pro.

VibeThinker-3B, based on Qwen2.5-Coder-3B, underwent a transformation using the Spectrum-to-Signal process for retraining: the model first accumulated a spectrum of problem-solving from easy to difficult tasks, then amplified correct solution signals through reinforcement learning. The training process utilized a 64K large thinking space to prevent interruption of the reasoning steps.

For mathematical and programming tasks, the model collected excellent self-generated steps for self-distillation and introduced step-level assessment during answering to perform self-checks. The self-check mechanism ultimately raised the AIME26 math test score from 94.3 to 97.1.

The development team proposed the "Parameter Compression-Coverage Hypothesis" in the report, suggesting that logical reasoning is a highly compressible ability, mainly relying on rules and error correction. A 3B small model can achieve top-tier performance, while open-domain knowledge requires a massive number of parameters for rote memorization. Due to its parameter size limitation, VibeThinker-3B's coverage of common-sense open knowledge still lags behind large models.

The team emphasized that the development goal is not to replace large models with small models but to explore the capability boundaries of compact models under a well-defined validation mechanism.