en.Wedoany.com Reported - Stepfun has launched its first large model native AI terminal brand, STEPX, along with the agent-native operating system Step AOS and the agent companion Step Amoo. It also unveiled the world's first large model native agent phone, the STEPX Neo.
Currently, most AI phones on the market primarily overlay AI capabilities on traditional operating systems. However, traditional systems are designed for manual operation, making it difficult for agents to directly invoke software functions and support cross-application long-process tasks. The STEPX Neo, equipped with Step Amoo and based on the Step AOS operating system, has achieved a transformation from "human-controlled phone systems" to "human-machine collaboration."
Stepfun has deployed a full-modal model matrix covering audio, image, video, and action, including the flagship Pro series, standard Flash series, and edge-side Edge series. The recently released Step Edge edge-side base model ranks first in its class across 29 authoritative benchmark evaluations.
Step AOS achieves millisecond-level environmental perception and voice interaction through a sensor matrix, and endows agents with screen semantic understanding capabilities. Based on its proprietary dual-domain six-layer structured memory and three-step continuous evolution technology, this operating system has reached SOTA levels in international evaluations such as PersonalMem, LongMemEval, and HaluMem. The system supports cloud-edge collaborative routing, achieving edge-side responses in hundreds of milliseconds and deep reasoning loops on the cloud side.
To address the memory wall, decision wall, and action wall issues of traditional operating systems, Stepfun built Step AOS from the ground up. At the computing level, the system's supply engine uniformly schedules heterogeneous computing power from CPU, GPU, and NPU, enabling full lifecycle management of edge-side models. At the data level, a unified semantic data layer processes perception, behavior, and personal data into semantic files. At the application and service level, an atomic capability engine decomposes system and application functions into thousands of minimal capability units, covering four major categories: communication, applications, files, and systems, all open via the MCP standard.
Specifically, the system's underlying architecture has been restructured at three levels: unified scheduling and optimization of heterogeneous computing power at the computing level; construction of a unified semantic data layer at the data level; and atomic decomposition of system functions at the application and service level. These designs provide agents with a dedicated operating environment.
The STEPX Neo, equipped with Step AOS and Step Amoo, has obtained the L3 level certification from the first batch of national standards for "Artificial Intelligence Terminal Intelligence Grading," making it the only agent phone to receive this level of certification. Stepfun STEPX has partnered with ecosystem players such as Ctrip, Alipay, Didi, Meituan, WPS, and CapCut, covering scenarios including travel, public services, local lifestyle, office productivity, and content creation.
At the press conference, STEPX, in collaboration with the Shanghai Artificial Intelligence Laboratory, released the "New Generation Agent System Security Technology White Paper" and the "Edge-Side Large Model Network Security Guide," proposing a "trustworthy, visible, controllable, and reversible" agent security framework. This framework requires all operations to be completed in a trusted execution environment, with steps auditable and traceable, behaviors isolated in a sandbox, and support for one-click revocation and step-by-step rollback.
