Mingshi Intelligence, Tsinghua University, and OpenBMB launched the edge-side multimodal model MiniCPM-V4.6. With only 1.3B parameters, it achieves top-tier performance among similar-sized models. Optimized for edge deployment, it reduces memory needs to 6GB and significantly boosts inference speed. The model leverages the LLaVA-UHD v4 technology to cut encoding overhead. It is open-source, supports major development frameworks, and is being deployed by partners like Lenovo and Xiaomi in automotive, smart home, and industrial applications.