STEP3-VL-10B is a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. Despite its compact 10B parameter footprint, STEP3-VL-10B excels in visual perception, complex reasoning, and human-centric alignment. It consistently outperforms models under the 10B scale and rivals or surpasses significantly larger open-weights models (10×–20× its size), such as GLM-4.6V (106B-A12B), Qwen3-VL-Thinking (235B-A22B), and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL.