en.Wedoany.com Reported - On July 9, StepFun will launch a new generation of agent terminal, providing a new terminal solution for the Agent era. Related product directions include AI terminal brands, agent systems, and AI agent phones, with the business focus extending from large model capabilities to hardware terminals and on-device agent entry points.
The technical core of the new generation of agent terminal is not just integrating large models into phones or hardware devices, but placing models, operating system interaction, on-device perception, task execution, and cloud capability scheduling in the same chain. Traditional phone assistants are mostly limited to voice wake-up, app invocation, and information queries, while the agent terminal needs to understand user intent, recognize screen and environmental context, break down multi-step tasks, and perform continuous operations across applications such as communication, scheduling, search, photography, documents, payments, travel, and office work. To achieve such capabilities, the terminal side requires multimodal input, on-device model inference, permission management, application interface invocation, and task state memory; the cloud side needs to provide larger model capabilities, complex reasoning, long-context processing, and cross-device synchronization. StepFun's entry into agent terminals focuses on cloud-device collaborative architecture, agent systems, GUI operation models, multimodal interaction, and phone hardware adaptation.
If AI agent phones enter the implementation stage, the hardware side needs to address more than just chips and screens. Microphone arrays, cameras, sensors, on-device computing power, system permissions, power control, privacy isolation, and local storage all affect whether the agent can run stably.
StepFun has previously identified "AI + Terminal" as an important direction, with product capabilities covering scenarios such as phones, automobiles, and multimodal interaction. Unlike simple model APIs, the agent terminal requires the model to truly enter users' high-frequency devices and execute actionable tasks in real applications. Phones are the most typical entry point because they simultaneously possess identity, payment, communication, camera, location, application ecosystem, and personal data; once the agent system is embedded in the phone, the model is no longer just answering questions but can perform cross-application operations, information organization, content generation, process execution, and device collaboration based on user instructions.
This terminal solution also involves collaboration between device manufacturing and large model companies. For AI agent phones to achieve batch deployment, model companies, ODM manufacturers, system software teams, application ecosystems, and chip platforms need to work together; simply pre-installing large models as applications on phones cannot create a true agent terminal experience. A real agent system requires establishing stable channels between underlying permissions, application interfaces, on-device inference, and cloud models, and continuously optimizing multi-step task execution capabilities through training.
Future focus areas for StepFun's new generation of agent terminal will center on several technical aspects: how the agent system invokes phone applications, how many local tasks the on-device model can handle, how multimodal input is converted into executable operations, how cloud large models and local systems divide work, how private data is isolated, and whether AI agent phones possess continuous task execution capabilities independent of traditional voice assistants.
