Clash of Strategies in AI Smartphones: Nubia, StepFun, and Honor at WAIC 2026

On July 17, 2026, during the World Artificial Intelligence Conference (WAIC) held in Shanghai, three AI agent smartphones were unveiled. Aside from the robotics booths, these devices were the highlight of the event, drawing the most visitors. The featured devices were the Nubia NaviX Ultra, StepFun STEPX Neo, and the Honor Robot Phone.

Just prior to the event, the Cyberspace Administration of China released a registry for “on-device generative AI services for smartphones.” The list included devices from Apple, Huawei, OPPO, vivo, Xiaomi, Samsung, and Nubia (ZTE), marking the first time regulatory authorities recognized on-device AI as an independent category. However, Honor and StepFun were absent from this registry.

According to industry insiders, 2026 is the first year when four key conditions—on-device computing power, model capabilities, regulatory clarity, and market demand—have been simultaneously met. Data from IDC indicates that global smartphone shipments declined by 6.7% year-on-year in Q2 2026 to 277.5 million units, marking the fifth consecutive quarter of decline. With the market seeking “new narratives to drive replacements,” AI agents are being viewed as a beacon of hope. WAIC provided the first comprehensive review platform for this competitive landscape.

This article examines the differing approaches of the three devices, their current positioning in the market, and the challenges faced by AI agent smartphones.

Strategic Differences Across Three Approaches

The three AI agent smartphones showcased at WAIC addressed challenges at different levels: the application layer, system layer, and perception layer. While Nubia’s approach has shown rapid short-term progress, it has also encountered significant hurdles early on.

Nubia’s first-generation model, the M153, utilized a system where an AI operated the graphical user interface (GUI). This approach garnered such high demand that 30,000 units sold out almost immediately. However, simulating screen taps conflicted with app profits and privacy concerns. Restrictions on AI operation rights ultimately made this approach unsustainable.

The second-generation Nubia NaviX Ultra took a different direction. By adopting the MCP (Model Context Protocol) and A2A (Agent-to-Agent) protocols, the AI embedded in the smartphone system now collaborates directly with agents inside individual apps at a foundational level, eliminating the need for screen-tap simulations. All core inferences are designed to be performed locally. Industry sources suggest the phone is equipped with a Qualcomm Snapdragon processor, a 6,000mAh battery, and supports 90W fast charging, highlighting the demands for processing power and energy efficiency in running local agent models.

However, the essence of AI handling tasks on behalf of users reduces the frequency and duration of app usage. This directly impacts click-through rates and advertising revenue on app launch screens, raising concerns about sustainability. The success of this approach depends heavily on whether a sustainable revenue-sharing model can be established among smartphone manufacturers, model developers, and internet platforms.

Rebuilding the System Layer

StepFun has taken a more foundational approach by building an independent operating system, “Step AOS,” from scratch. Just before the WAIC on July 13, StepFun unveiled its Step AOS agent-native OS, the built-in agent StepFun Amoo, and its first smartphone, the STEPX Neo. Step AOS standardizes chip computing resources, organizes scattered user data into formats readable by AI, and breaks down smartphone functionalities—such as calls and app launches—into the smallest possible units. Agents can then freely assemble and invoke these functions like building blocks.

Yin Qi, the chairman of StepFun, explained, “Opening doors in old systems for agents makes them visitors at best. Building a house for agents allows them to truly become natives.” While reconstructing the system at the foundational level addresses issues of permissions and task scheduling, building a supportive partner ecosystem will take time. Notably, StepFun’s initial ecosystem partner list does not include WeChat, although Yin Qi has stated that deep discussions are ongoing with Tencent.

The STEPX Neo is expected to launch in late October. Through standardized protocols, Step AOS plans to partner with platforms like CTrip, Alipay, Didi, and Meituan. While tasks such as searching, comparing, and booking can be fully handled by agents, user confirmation will still be required for payments. Rebuilding the system layer offers the potential to fundamentally eliminate barriers between apps but will require time and negotiation for the ecosystem to mature.

Expanding Perception and Physical Form

Honor’s most striking difference lies in its device design. The company’s Robot Phone features a retractable four-degree-of-freedom titanium alloy mechanical gimbal, capable of extending in 0.8 seconds and supporting 360-degree object tracking. While the other two companies focus on enabling their agents to operate in the digital realm, Honor aims to extend its agent’s capabilities to perceive and interact with the physical world.

Honor’s approach hinges on multimodal perception. Input signals including voice, gestures, gaze, and movements are utilized, with the mechanical gimbal serving as a carrier for sensors. The smartphone is equipped with capabilities for spatial positioning and object tracking. This is powered by the Magic Agent large-scale model matrix, co-developed by Honor and Alibaba, which integrates a general-purpose large-scale model for understanding user intent with multiple specialized models for handling images, voice, and actions.

Honor’s CEO, Li Jian, has confirmed that the product is ready, with industry sources suggesting an August launch. As the device manufacturer, Honor has complete control over hardware design, sensor configuration, and system scheduling—an advantage not shared by pure model companies. However, commercial validation of innovative physical forms often requires a longer cycle, and the cost, yield, and durability of the four-degree-of-freedom gimbal will need to withstand market scrutiny.

Evaluation Criteria for AI Smartphones

Whether smartphones can genuinely perform tasks on behalf of users depends on several factors. Nubia’s president, Ni Fei, summarized these criteria as “comprehension, execution, memory, and security.” These four benchmarks provide a framework for evaluating the current generation of AI smartphones.

To achieve “comprehension,” multimodal input from vision and audio is essential. During StepFun’s presentation, an example of photo editing was shown. Traditionally, users would have to click through steps like “smooth skin” or “enlarge eyes,” but with the new method, users simply open a photo, circle the area they’re dissatisfied with, and say, “Make the eyes bigger.” The action of circling is captured by the visual module, while the voice command is parsed by the language model, and both signals are fused within the system to execute the intended task.

“Execution,” on the other hand, depends on how well the industry collaborates. StepFun, for example, has broken down functions like ticket booking, navigation, and payments into callable interfaces. By introducing the A2A protocol in collaboration with multiple smartphone makers, WeChat enables AI assistants to directly access app functionalities. Ideally, a user could say, “Plan a trip to visit clients in Shanghai next week without waking up early,” and the agent would read the calendar, check the weather, book flights and hotels, and handle everything in the background.

However, even a 90% success rate for each step in a five-step cross-app task would result in an overall success rate of just 59%. One major cause of failure lies in the inability to remember user preferences and context, a critical distinction between “true agents” and “pseudo-agents.”

The three approaches are tackling challenges at different levels. Nubia eliminates barriers at the application layer, StepFun reconstructs scheduling at the system layer, and Honor expands perceptions to the physical realm. However, a comprehensive AI agent smartphone experience will require all three layers to function together. The race to establish dominance in this category will depend on who can solidify their strengths first and then complement the remaining layers.

As the industry looks to AI agents to provide the “new narrative” needed to drive smartphone replacement cycles, WAIC has served as the first comprehensive review platform for this burgeoning competition.

Editorial Opinion

In the short term, Nubia’s adoption of the MCP/A2A protocol appears to be the most feasible, likely leading many Android smartphone manufacturers to follow suit between late 2026 and 2027. However, unless the revenue-sharing problem with app platforms is resolved, the user experience will remain limited. StepFun’s system-level reconstruction provides a more fundamental solution but requires more time to build its ecosystem, limiting its adoption until at least early 2027.

In the long term, Honor’s innovation with mechanical gimbals may have the most significant impact. As digital-only agents become less differentiable, the extension of perception into the physical world could give rise to a new category, merging smartphones with robotics. However, overcoming challenges related to cost and durability will be crucial, and the market will need one to two years to render a verdict.

From the editorial perspective, the biggest shared challenge for all three approaches is the “success rate of multi-step tasks,” which remains the primary bottleneck for AI agent smartphones.

References

• “努比亚、阶跃、荣耀，谁在定义AI手机？”, by 定焦One — 钛媒体, 2026-07-19T08:31:26.000Z (ARR)
• Source URL: https://www.tmtpost.com/8070760.html