Recently, the THUNLP Lab of Tsinghua University and Mianbi Intelligence jointly launched an innovative client-side GUI Agent called **AgentCPM-GUI**, bringing a new breakthrough in human-computer interaction on mobile devices. This intelligent agent is built based on the **MiniCPM-V** model with only **8B** parameters. It takes the screen image of the phone as input, supports both Chinese and English operations, and can automatically execute tasks proposed by users, demonstrating strong GUI element positioning capabilities.
AgentCPM-GUI covers more than **30 mainstream Chinese apps**, including **AutoNavi Maps**, **Dianping**, **Bilibili**, and **Xiaohongshu**, capable of accurately recognizing and operating APP interface elements to meet diverse user needs. Whether it's navigation, ordering food, or content browsing, AgentCPM-GUI can efficiently complete tasks, greatly enhancing the user experience.
Notably, this model enhances planning reasoning ability through **RFT (Reasoning Before Thinking)** technology. Before executing user instructions, AgentCPM-GUI will first reason and think, generating more accurate action sequences to improve the success rate and reliability of task execution. The application of this technology makes its performance particularly outstanding in the client-side AI field.
As a lightweight and high-performance model, AgentCPM-GUI runs smoothly on mobile devices, showcasing the deep strength of Tsinghua University's THUNLP Lab and Mianbi Intelligence in AI technology. In the future, this GUI Agent is expected to further promote the popularization and application of client-side AI, helping smart devices move toward a more efficient interaction era.