When Will AI Terminals Reach Their ‘iPhone Moment’? Insights from WAIC

Parameter paradox and WAIC showcase

Over the past two years the AI industry has shown a growing gap between rapidly improving large‑model capabilities—such as inference speed, long‑context handling, multimodal reasoning, and tool calling—and ordinary users’ perception of AI, which remains limited to isolated chat windows. The author calls this the “parameter paradox”: stronger models but weaker user experience. At WAIC on July 13, Step Star (阶跃星辰) presented a solution to this paradox.

STEPX Neo and the world’s first model‑native AI terminal

Step Star unveiled the STEPX Neo, marketed as the world’s first large‑model‑native AI terminal (brand STEPX) running the native agent operating system Step AOS and the personal agent Amoo. The device received the first L3‑level certification in the national “Artificial‑Intelligence Terminal Intelligence Grading” series, the highest open‑test tier for AI terminals.

During the WAIC demo the author asked the phone to locate a charging station and deliver a drink. Unlike traditional GUI agents that click through screens, STEPX Neo directly invoked services from Alipay and other apps, turning a multi‑click workflow into a structured service call and eliminating instability caused by UI changes.

Contrast with the conventional “OS + AI” model

Most vendors adopt an “OS + AI” strategy, layering a generative model on top of an existing Android or Linux system (e.g., Apple Intelligence adds models to Siri and system apps). This approach provides limited integration; the model remains a guest that must be called explicitly, and cross‑app coordination is weak. In contrast, Step AOS is built from the ground up for agents, treating the agent as a native resident rather than a visitor.

Step AOS architecture

Step AOS reorganizes resources into three layers: new facilities, new capabilities, and new interaction patterns.

Computation layer : a unified pool schedules CPU, GPU, NPU and other heterogeneous resources. Latency‑sensitive tasks run on‑device; complex, multi‑step planning can fall back to cloud models, creating a seamless “fast‑reaction” + “deep‑thinking” hybrid.

Data layer : a unified semantic data tier converts perception, user behavior, and personal data into agent‑readable semantics, providing contextual continuity across time, people, and tasks.

Application & service layer : an atomic‑capability engine breaks system functions and external services into composable units. Partnerships with companies such as Ctrip, Alipay, Didi, and Meituan use protocol interfaces instead of GUI simulation, giving clear data boundaries, authorization, and feedback.

Memory, decision and security capabilities

Step AOS implements a “dual‑domain three‑step” memory structure. The user domain captures identity, preferences, and context; the agent domain stores its own knowledge. The three steps—record, organize, recall—operate in the background, achieving SOTA results on PersonaMem and LongMemEval and a 15 ms recall latency for everyday Q&A.

Decision‑making relies on the on‑device Step Edge model for quick, privacy‑sensitive tasks (e.g., setting alarms, fetching photos) while delegating heavy reasoning to cloud flagship models. Local inference completes in sub‑hundred‑millisecond latency with >99 % success, and Step Edge ranks first among 29 on‑device benchmarks.

Security is addressed through a four‑dimensional framework: trust, visibility, controllability, and reversibility. Actions execute in a trusted environment, steps are auditable, permissions are granted on‑demand and revoked after use, and erroneous operations can be rolled back. The system’s safety guidelines are documented in the “Intelligent‑Agent System Security Whitepaper” co‑authored with the Shanghai AI Lab.

Result‑oriented interaction paradigm

Step AOS enables a shift from “process interaction” to “result interaction”. Users state an intent, and the agent orchestrates planning, service calls, and confirmation, delivering a final outcome while hiding low‑risk steps and prompting for high‑risk confirmations. This contrasts with traditional apps that expose fixed pages and require users to manually sequence functions.

The author cites a formula presented at WAIC: Agent ability = Model ability × Agentic OS ability. If either factor is weak, model intelligence cannot reliably translate into usable capability.

Demo scenarios include automatic photo organization after an exhibition, context‑aware ticket booking from a concert poster, and a workplace assistant that calculates travel distance, costs, and dispatches a courier without repeated prompts.

Implications

The STEPX Neo exemplifies how integrating model, system, and hardware can create an AI terminal that delivers a consumer‑friendly “iPhone moment” for intelligent agents, potentially setting a new standard for AI‑driven personal devices.