Yin Qi Fires the First Shot for Global Agent Smartphones

Two days before the opening of the World Artificial Intelligence Conference, the consumer tech sector is buzzing with activity. Right after Nubia announced plans to launch an agent-powered smartphone at WAIC, StepNova held a pre-emptive launch event in Shanghai, unveiling its full-stack intelligent agent terminal solution:

The launch introduced STEPX, a native large-model AI terminal brand, StepAOS, an agent-native operating system, Amoo, a personal intelligent agent, alongside the debut agent smartphone STEPX Neo.

Over the past two years, nearly all AI-branded phones on the market have followed the same shortcut: leaving the underlying system untouched, adding a standalone AI assistant window with basic text generation and image processing features, essentially slapping an AI skin onto traditional smartphones. Few players were willing to revamp the core operating system, as reconstruction requires massive R&D investment and incurs huge costs for ecosystem adaptation. Phone manufacturers avoided this path, while most large model companies remained content with delivering cloud-based software services.

This time, however, Li Ying and StepNova have broken this unspoken industry consensus. Their approach combines the dedicated terminal brand STEPX, the Step AOS operating system fully reconstructed for intelligent agents, the personal agent Amoo capable of autonomously planning and executing tasks, and the world's first large-model native agent smartphone STEPX Neo.

As an entrepreneur with deep roots in computer vision who has witnessed the full lifecycle of AI commercialization, Ying believes that relying solely on cloud-based large models and add-on AI features can never truly reshape the interaction logic between humans and mobile devices. Only by embedding intelligent agents deep into the system kernel, dismantling all application capabilities to enable cross-app autonomous scheduling, can portable terminals truly gain genuine autonomous execution intelligence.

Native Agent Systems Are Fundamentally Different from Add-on AI Features

There is no shortage of products labeled as AI phones on the market, but the vast majority follow a uniform logic: adding a floating window or a dedicated AI entry panel on top of Android or iOS, which can generate summaries, draft copy, and recognize visual content — essentially remaining an independent app-level feature. When users need to complete cross-application tasks, they still have to manually switch between apps, with the AI only capable of launching the target app at best, leaving all subsequent operations to the user.

StepAOS takes a completely different path. The official description states it redesigned the underlying architecture around intelligent agents — a concept that sounds abstract, but delivers strikingly tangible differences in real-world use.

This system breaks down all apps, files, and system capabilities into invocable atomic functions, dismantling standalone apps into granular functional modules stored in a unified system capability pool. With system-level permissions, the intelligent agent can directly schedule corresponding functions from the pool after receiving user instructions, automatically chaining them together per task workflows, without requiring users to manually launch apps, fill in information, or confirm payments.

For example, to arrange a short trip, users only need to state their destination and time preferences, and the agent can automatically call Ctrip to book tickets and hotels, Meituan to find local restaurants, Didi to hail rides, and Alipay to complete payments — the entire process never requires switching to any individual app interface. The same logic applies to office scenarios, where organizing documents, generating presentations, and editing footage can be completed by the agent directly invoking corresponding capabilities from WPS and CapCut.

This experience relies on a full reconstruction of the system kernel. Traditional operating systems are designed around the logic of "humans operating apps", with computing power, storage, and perception capabilities all allocated to foreground applications. StepAOS, by contrast, redistributes resources around the operational logic of intelligent agents, building a heterogeneous computing unified scheduling pool where CPU, GPU, and NPU can dynamically and flexibly allocate resources to meet real-time inference demands of on-device models. At the data layer, a unified semantic data layer standardizes multimodal perception and user behavior data, paired with a high-speed hybrid database to achieve millisecond-level data retrieval.

Memory capability is another core differentiator. StepAOS adopts a dual-domain three-step memory structure, where the user domain accumulates personal habits and scenario preferences, and the agent domain stores task execution experience and domain-specific knowledge. Users no longer need to repeatedly state their dietary preferences or travel habits, as the agent can make decisions directly based on historical memory.

This smartphone has passed the L3-level test under the national standard "Intelligent Classification of Artificial Intelligence Terminals". This newly released national standard divides terminal intelligence levels from low to high into four tiers. The L3 assistant level requires terminals to fully understand complex intentions, proactively clarify ambiguous instructions, independently break down tasks and plan workflows, flexibly invoke multiple tools, and possess long-short term memory and multimodal generation capabilities.

Prior to this, the vast majority of AI phones on the market remained at the L2 tool level, only capable of executing single or very few steps under explicit instructions.

Naturally, the device-cloud collaborative architecture is also integrated. Simple local operations are processed by the on-device StepEdge model, delivering responses within hundreds of milliseconds to save bandwidth and protect privacy. Complex multi-step planning and deep content generation tasks are automatically offloaded to cloud-based large models. The system dynamically selects the optimal model based on task complexity, privacy level, and network conditions, escalating processing tiers when necessary to balance speed and output quality.

The hardware modifications are not overly radical: a secondary interactive display on the back acts as a dedicated persistent window for the intelligent agent. Users can view the agent's task progress and notification alerts without lighting up the entire main screen, and perform basic interactions directly through the secondary display. This design philosophy frames the agent as a continuously running background service that does not demand the user's full attention, with the secondary display perfectly catering to this lightweight, always-on interaction requirement.

Large Model Companies Venturing into Hardware Is Both a Necessary Response and an Inevitable Trend

Why would a dedicated large model company dive into smartphone manufacturing? For StepNova and Li Ying, this move is driven both by industry pressures and inherent strategic necessity.

Founded in 2023, StepNova entered the market at the tail end of the "Hundred Models Battle", officially releasing its hundred-billion-parameter model in 2024, making it a relatively late entrant among the first-tier large model players. Over the past two years, the C-end large model landscape has rapidly solidified, with leading products leveraging first-mover advantages and scenario positioning to accumulate massive user bases and entrenched usage habits.

Although StepNova has invested heavily in the multimodal domain, earning itself the nickname "Multimodal Overachiever" in the industry, it has never launched a breakout C-end consumer app. After completing its Series B funding round at the end of 2024, the company explicitly announced a strategic shift, scaling back its C-end operations to focus on the "AI + Terminal" paradigm, prioritizing four core scenarios: automotive, smartphones, embodied intelligence, and IoT.

In early 2026, Li Ying officially took office as Chairman, bringing in over 5 billion yuan in Series B+ funding that set a new record for the largest single financing round in China's large model sector at the time. This funding gave StepNova the capital cushion to develop hardware, and more critically, Ying brought his vision of a full software-hardware closed loop into the company.

Those familiar with Ying know that he went through the complete AI commercialization lifecycle at Megvii. Starting from facial recognition technology, to the setback of its IPO attempt, over a decade of entrepreneurial experience gave him a pragmatic perspective on technology implementation: pure algorithm companies can hardly build a sustainable commercial closed loop, and all achievements without end-to-end integration are only temporary. This judgment largely forms the underlying logic behind StepNova's pivot to terminals.

At the current stage of the large model industry, competition at the pure model layer has become increasingly fierce, with severe homogenization and persistently unclear commercialization paths. ToB services face high customization costs and scaling barriers, while ToC products suffer from low user willingness to pay and exorbitant traffic acquisition costs. Terminals represent the optimal value carrier: embedding model capabilities directly into hardware not only creates differentiated product experiences, but also paves a more stable path to commercialization.

Why smartphones specifically? The answer is straightforward: smartphones are currently the most widely adopted, frequently used, and highly sticky personal intelligent terminals, making them the ideal portable carrier for intelligent agents. While smart glasses and humanoid robots hold long-term potential, they face extended implementation cycles and require significant user adoption cultivation. The smartphone market, despite intense competition, benefits from a mature industrial chain and well-established user habits, making it the fastest scenario for intelligent agents to launch and validate their value.

Among the three core goals Ying set for the company at the start of the year was launching an AI-native innovative hardware product within 12 to 15 months. The STEPX Neo launch came just half a year after the goal was established, moving faster than expected. This further confirms that StepNova's terminal strategy is not an impulsive decision, but a long-planned core layout.

Of course, large model companies have inherent weaknesses in hardware development. Supply chain management, mass production capabilities, and distribution channel deployment are competencies that traditional phone manufacturers have cultivated over decades, which cannot be rapidly bridged by capital and algorithm advantages alone. The product has not yet announced specific configurations, pricing, or release dates, only confirming it will be publicly demonstrated at the World Artificial Intelligence Conference — a reflection of StepNova's cautious approach: showcasing the technical solution and product form first, gathering market feedback, then gradually advancing the mass production roadmap.

Interestingly, nearly in parallel with StepNova, Nubia also announced it will launch an agent-powered smartphone at WAIC, also emphasizing native intelligent agent experiences. On the side of established phone manufacturers, Huawei, Honor, Xiaomi, and vivo have all intensively upgraded their AI Agent capabilities this year, promoting cross-app task execution at the system level.

According to a Counterpoint report, smartphones with generative AI capabilities are projected to account for 45% of global shipments by 2026, with intelligent agent functionalities emerging as the core competitive differentiator in mid-to-high-end smartphones.

The entire industry recognizes that the next era of smartphones will be the era of intelligent agents. No one expected, however, that the first company to deliver a fully realized native agent smartphone would not be a legacy phone manufacturer, but a large model company.

In the Agent Era, Apps Will Not Disappear — They Will Evolve

During the launch event, Li Ying shared a thought-provoking perspective: In the intelligent agent terminal era, applications will not vanish, but will continue to exist through different carriers.

This statement directly addresses the long-running industry debate: will apps become obsolete after intelligent agents are widely adopted? Many argue that once agents handle all tasks, users will no longer need to launch individual apps, rendering applications irrelevant. The reality, however, is far more nuanced.

Intelligent agents essentially execute tasks by invoking capabilities from various services. Flight bookings rely on inventory from airlines and OTAs, food delivery draws on merchant resources from local lifestyle platforms, and office work leverages functional modules from professional software. These underlying service capabilities were originally delivered through apps. The emergence of intelligent agents does not change the services themselves, but transforms how users access them.

Previously, the user journey to access services was: locate the app icon, tap to launch, find the corresponding function in the UI, fill in information, and complete the operation. In the agent era, users only need to state their requirements, and the agent will automatically schedule corresponding service capabilities in the background, with users never needing to interact with any app interface.

For users, the entry-point attribute of apps disappears; but for service providers, their core capabilities are only invoked through a new mechanism, not eliminated.

The atomic capability engine in StepAOS is built exactly for this purpose: breaking down core functionalities of various apps into standardized atomic services, integrating them into a unified scheduling framework open for intelligent agents to invoke. StepNova has already established partnerships with platforms including Ctrip, Alipay, Didi, Meituan, WPS, CapCut, Baidu, and JD, covering high-frequency scenarios such as travel, local lifestyle, office work, and content creation.

The willingness of these leading platforms to integrate essentially stems from their recognition of intelligent agents as a promising new traffic entry point.

Nevertheless, significant challenges remain. While leading platforms can negotiate partnerships, what happens to the massive ecosystem of small and medium-sized developers? Under the traditional mobile internet ecosystem governed by app store distribution rules, developers build apps, list them on marketplaces, and users download and use them.

In the intelligent agent era, the entry point shifts to the agent, and services become atomic capability modules, completely transforming how developers create, distribute, and monetize their products.

Who will establish the rules for this new ecosystem, and how will benefits be distributed? These are the critical questions the industry must resolve moving forward. If all traffic is controlled by intelligent agents, will service providers lose bargaining power? Will developers shift from building products for end-users to building interfaces for agents? No clear answers exist for these questions yet.

Another unavoidable topic is security and privacy. Intelligent agents hold system-level permissions that can access sensitive functionalities such as payments, contacts, and location data. Any misoperation or unauthorized privilege abuse could lead to far more severe consequences than those from regular apps. StepAOS has implemented four layers of security protection, emphasizing trustworthiness, visibility, controllability, and traceability, allowing users to view the agent's full operation history at any time and revoke permissions instantly. However, building sustained user trust will require long-term validation through real-world usage.

From a broader perspective, the emergence of agent-powered smartphones is essentially redefining what a smartphone is. Over the past decade, the smartphone served as the primary entry point to the mobile internet, a container for individual apps. In the future, smartphones will evolve into portable carriers for personal intelligent agents, where user-device interaction no longer revolves around tapping icons, but conversing with a smart companion that understands personal habits and proactively assists with tasks.

This transformation will not happen overnight. The current STEPX Neo remains an early-generation product, and its real-world performance and ability to solve practical use cases will ultimately be judged by users after mass production. But its launch acts as a clear industry milestone, signaling to the entire ecosystem: the next phase of AI smartphone development has officially begun.

This article originates from the WeChat Public Account "Xin Shi" (ID: xinmouls), authored by Li Xiaodong, published with authorization from 36Kr.