Sina Weibo Open-Sources VibeThinker-3B, 3B-Parameter Model Matches Frontier Reasoning Performance

According to Beating intelligence, Sina Weibo team recently open-sourced VibeThinker-3B, a 3-billion-parameter reasoning model that achieves frontier-level performance on mathematics and coding tasks, with some metrics approaching or exceeding DeepSeek V3.2, GLM-5, and Gemini 3 Pro.

Built on Qwen2.5-Coder-3B, the model uses a Spectrum-to-Signal training pipeline with 64K reasoning space. Through self-distillation and step-level verification during inference, VibeThinker-3B improved AIME26 math test scores from 94.3 to 97.1.