MobiAgent - Shanghai Jiaotong University open source mobile intelligent body full-stack building framework

What is MobiAgent

MobiAgent is an open source mobile intelligent body toolchain from IPADS Lab of Shanghai Jiaotong University, which helps users to build their own mobile intelligent assistants. MobiAgent can be trained to understand natural language commands by recording the user's operation trajectory and generating high-quality data. The core features include efficient training process, unique "latent memory gas pedal" and innovative AgentRR acceleration framework, which significantly improves the efficiency of task execution.MobiAgent's architecture consists of three parts: planner, decision maker and executor, which are responsible for task planning, decision making and specific operation respectively. MobiAgent's architecture consists of three parts: planner, decision maker, and executor, which are responsible for task planning, decision making, and concrete operation respectively. It outperforms several well-known closed-source models in terms of task completion quality, and provides developers with full-process support from data collection to model deployment through the open-source project, which promotes the development of mobile intelligent body technology.

Features of MobiAgent

• Data collection: It can record the trajectory of the user's operation on the cell phone and provide a data base for subsequent training.
• intelligent training (religion): Generate high-quality training data and train proprietary intelligences using the collected data and generic VLM models.
• Mission planning and decision-making: With the Planner and Decision Maker modules, the intelligences understand natural language commands and make rational decisions.
• Efficient implementation: The Executor module is responsible for performing specific operations to ensure the successful completion of tasks.
• Accelerating repetitive tasks: Dramatically increase the speed of execution of repetitive tasks with the Subliminal Memory Accelerator and AgentRR Acceleration Framework.
• Model deployment: Supports deployment of trained smart body models to cell phones for easy access at any time.

Core Benefits of MobiAgent

• Efficient data collection and processingThe VLM model is a lightweight tool that records user trajectories and generates high-quality training data with a generalized VLM model, providing a solid foundation for intelligent body training.
• Strong mandate implementation capabilities: In real-world application scenarios, MobiAgent outperforms several well-known closed-source macromodels in terms of task completion quality, and more accurately understands and executes user commands.
• Significant performance gainsThe unique Latent Memory Accelerator and AgentRR Acceleration Framework dramatically improves the execution efficiency of repetitive tasks by 2 to 3 times, with an action reuse rate of up to 85%.
• Complete full-stack toolchainMobiAgent provides a complete solution from data collection, model training to final deployment, lowering the threshold for users to build a mobile agent from scratch.
• Open Source and Scalability: The project is open source, users can customize and expand according to their own needs, to promote the further development and application of technology.

What is the official website of MobiAgent

• paper address:: https://arxiv.org/pdf/2509.00531
• Github repository::https://github.com/IPADS-SAI/MobiAgent
• HuggingFace Model Library:: https://huggingface.co/collections/IPADS-SAI/mobimind-68b2aad150ccafd9d9e10e4d

Who MobiAgent is for

• average cell phone user: You want to use smart assistants to do everyday phone operations more efficiently, such as auto-replying to messages, quickly finding information, and so on.
• technology enthusiast: Interested in Artificial Intelligence and Mobile Intelligent Body technologies and want to explore and practice how to build and optimize their own mobile intelligent assistants.
• developers: Have a technical background and want to expand your business or research direction by developing more complex and personalized mobile intelligence applications through MobiAgent's open source toolchain.
• research worker: Scholars specializing in artificial intelligence, natural language processing, or mobile computing can use MobiAgent to conduct research and experiments that advance the state of the art.