PhoneBuddy - Tencent Hunyuan's open-source 4B parameter mobile agent model.

What is PhoneBuddy?

PhoneBuddy is an open-source 4B parameter mobile agent model from Tencent's Hunyuan team. Its core research focuses on training usable AI agents in real-world mobile scenarios. The model employs a hybrid RL training scheme combining real apps and mock apps. In 150 real-device evaluation tasks, its success rate surpasses GPT-5.4 for both single-app and WeChat mini-program tasks, and reaches 83.2% on AndroidWorld, demonstrating that this small model can be highly competitive through environment design optimization.

PhoneBuddy's main functions

• Understanding and operating mobile GUIIt can recognize screenshots of mobile phone screens and predict the next action such as clicking, swiping, and typing.
• Single App Task ExecutionComplete operations such as searching, settings, and content creation within a single native application.
• Cross-App information flowIt enables the transfer of information between multiple applications to complete complex tasks that require collaboration.
• WeChat Mini Program OperationIt supports tasks such as searching, booking, and querying within the WeChat Mini Program ecosystem.
• Real-world task verificationVerify whether the task has been truly completed, such as whether the message was sent or the document was saved.

The technical principles of PhoneBuddy

• Two-stage training architectureBy using Shared SFT, the model learns the operation trajectories of both real and mock apps simultaneously, establishing a unified basic capability for mobile phone operation. In the RL stage, the effects of training only in a real environment and training with a combination of real and mock apps are compared.
• Real+Mock Hybrid RLThe Real App provides real business logic, account status, and side effects; the PhoneWorld Mock App provides resettable, verifiable, and scalable training signals. The two complement each other to resolve the contradiction between real but difficult-to-train and controllable but not realistic.
• PhoneWorld Environment SetupReconstruct a working Android Mock App from the actual GUI usage structure, preserving page structure, navigation relationships, interactive elements, and task validators, enabling RL to obtain stable reward signals.
• Task-level verification mechanismDirectly check the final status of the task, such as whether the hotel budget has been filtered and whether the leave request is pasted correctly, to ensure that the agent truly achieves the user's goal.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use PhoneBuddy

• Environmental preparationDownload and configure the PhoneBuddy-4B model weights and dependencies, and prepare an Android real device or emulator as the execution environment.
• Data collectionCollect real app operation trajectories for SFT stage training, and build or connect to PhoneWorld Mock App to obtain verifiable interaction data.
• Model trainingFirst, perform Shared SFT training to enable the model to master a unified operating format, and then select a Real-only or Real+Mock environment for RL fine-tuning.
• Task DeploymentThe trained model is connected to the mobile execution layer (such as PhoneHarness), and task instructions are issued through GUI or CLI.
• Result verification:Use the task validator to check the final execution status to confirm that the task has actually been completed and that the page does not just look correct.

PhoneBuddy's core advantages

• Real-world scenario orientationTraining is directly geared towards real mobile phones and real apps, rather than just optimizing offline benchmark scores.
• Small model, high performanceThe 4B parameter alone surpasses GPT-5.4 in single App and WeChat Mini Program tasks, proving that environment design is more important than simply stacking parameters.
• Verifiable RL trainingThe PhoneWorld Mock environment provides a stable reward, enabling RL training to scale scalably.
• Full technology stack open sourceEverything from environment, training, execution to security and privacy assessments is made public.
• Task completion guaranteeBy constraining follow-up and verifying information transfer, we ensure that the Agent truly executes the user's intent.

PhoneBuddy project address

• Project official websitehttps://phonebuddyai.github.io/
• GitHub repository: https://github.com/PhoneBuddyAI/phonebuddy
• HuggingFace model libraryhttps://huggingface.co/PhoneBuddyAI/PhoneBuddy-4B
• Technical Papers: https://phonebuddyai.github.io/assets/paper.pdf

PhoneBuddy's Competitive Product Comparison

Application scenarios of PhoneBuddy