MiniCPM-V 4.7, the new generation of edge-side multimodal large model from ModelBest, was officially released on October 13. Its 35B-A3B version adopts a Mixture-of-Experts (MoE) architecture with a total of 35 billion parameters, activating approximately 3 billion parameters per inference. ModelBest stated that this is the first open-source MoE model capable of running smoothly and losslessly on edge devices, requiring only about 3GB of RAM to achieve a decoding speed of over 15 tokens per second on an iPad Pro. The model enhances edge-side inference speed by approximately 5 times through its Look-Ahead Expert Scheduling (LES) technology and supports iOS, Android, and HarmonyOS, adapting to chip platforms such as Rockchip, Intel, and Apple.