May 13, 11:52 AM

OpenBMB Releases AgentCPM-GUI Built on MiniCPM-V With 8B Parameters, Supporting 30+ Chinese Apps and RFT-Enhanced Reasoning

OpenBMB and collaborators, including Tsinghua University and Mianbi Intelligence, have released AgentCPM-GUI, an open-source on-device graphical user interface (GUI) agent designed to operate Android applications autonomously. Built on the MiniCPM-V model with 8 billion parameters, AgentCPM-GUI processes smartphone screenshots as input to execute user-specified tasks efficiently. It is the first open-source GUI agent fine-tuned specifically for Chinese apps, supporting over 30 mainstream applications such as Gaode Map, Dazhong Dianping, Bilibili, and Xiaohongshu. The agent features strong bilingual grounding based on an Android dataset and employs reinforcement fine-tuning (RFT) to enhance planning and reasoning capabilities. Additionally, it optimizes task execution by compressing the average action length to 9.7 tokens, improving inference efficiency on mobile devices. The development emphasizes on-device operation for greater trustworthiness compared to remote servers.

#Open #Tsinghua University #Mianbi Intelligence #Agent #Android #Chinese #Gaode Map #Dazhong Dianping #Bilibili #Xiaohongshu

Written with ChatGPT (GPT-4).

Sources

Additional media

Image #1 for story openbmb-releases-agentcpm-gui-built-on-minicpm-v-8b-parameters-supporting-30-rft-187404d7

Image #2 for story openbmb-releases-agentcpm-gui-built-on-minicpm-v-8b-parameters-supporting-30-rft-187404d7

Image #3 for story openbmb-releases-agentcpm-gui-built-on-minicpm-v-8b-parameters-supporting-30-rft-187404d7

OpenBMB Releases AgentCPM-GUI Built on MiniCPM-V With 8B Parameters, Supporting 30+ Chinese Apps and RFT-Enhanced Reasoning

Sources

Additional media

Similar Stories