VibeThinker-1.5B is a small dense language model designed to challenge the notion that small models lack robust reasoning capabilities. The model uses the Spectrum-to-Signal Principle (SSP) as its core post-training methodology to elicit strong reasoning performance. It demonstrates competitive results on mathematical benchmarks (AIME24, AIME25, HMMT25) and coding tasks, surpassing initial DeepSeek R1 across multiple evaluations while being 1.5B parameters versus R1’s 671B. Model weights and technical reports are publicly available on Hugging Face.