Mobile phone manufacturers are getting closer to Doubao.

There are still 30,000 units of the Nubia M153 left unsold. Doubao's cooperation map on the mobile side is expected to expand further.

Recently, Jiemian News reported that ByteDance is promoting AI mobile phone cooperation with hardware manufacturers such as vivo, Lenovo, and Transsion, and pre - installing AIGC plugins on their devices. Multiple vivo employees confirmed to the media that the two parties have confirmed the cooperation and are discussing details.

However, based on the vague statement of this news alone, it is still impossible to confirm the specific nature of these "cooperations" and whether they are equivalent to the previous cooperation plan between ByteDance and ZTE. Previously, on the nubia M153, ByteDance deeply built a GUI Agent at the mobile phone OS layer, which could automatically take over the user - specified cross - app operations. However, it subsequently faced "countermeasures" from apps such as WeChat and Taobao.

But it is not difficult to see from the latest cooperation rumors that the attempt with the M153 has still attracted the attention of the mobile phone industry. As ByteDance enters the market with Doubao's large - model capabilities, the native AI ecosystem of terminal manufacturers is accelerating its expansion.

This trend can be glimpsed at the recent Volcengine FORCE Prime Conference.

Many mobile phone manufacturers, including ZTE, Honor, Xiaomi, and vivo, "stood up" for Volcengine. However, except for the M153 brought by Nubia, most mobile phone manufacturers' relationship with Doubao is mainly about invoking model capabilities. The product form, interaction mode, and system - level integration at the Agent level are still mainly led by the mobile phone manufacturers.

During the group interview session of this Prime Conference, Tan Dai, the president of Volcengine, said in an interview with Facing AI, "The value of AI lies in expanding users' needs in a more convenient way."

Obviously, regardless of how ByteDance's cooperation with these manufacturers unfolds in the future, an industry trend has become increasingly clear. The potential of the system - level Agent is huge. Even though there are some obstacles to its implementation at present, the prototype of the future can already be seen on the M153.

Whether through cooperation or self - research, large - model manufacturers and mobile phone manufacturers have shouted the slogan of "growing and strengthening" and are accelerating into this track.

1

The reason why mobile phone manufacturers are accelerating their embrace of large - model manufacturers like ByteDance is that there is a naturally large and active user base in the field of native mobile phone assistants.

According to the Q3 2025 report of QuestMobile, as of September 2025, the MAU of mobile AI applications reached 729 million. Among them, the scale of "mobile phone manufacturers' AI assistants" was about 535 million, "native AI apps" were about 287 million, and "In - App AI" was about 706 million.

In other words, in different forms of native AI, there is an obvious "hierarchical difference" in users between terminals and apps: native AI apps rely on users to actively click to trigger, while system - level assistants are naturally in the front position of "ready - to - use" and can maintain daily active users without users' deliberate clicks.

In fact, the exploration of "how to activate AI at the system level" and "how to make the assistant not just stay at voice Q&A" has almost become a consensus among mainstream domestic mobile phone manufacturers. However, limited by model capabilities and engineering stability, most solutions have finally stayed at the stage of "enhanced voice assistants", and the emergence of the Doubao mobile phone assistant has directly kicked open this door.

Although the effect presented by the nubia M153 is still controversial in the industry, after the Doubao mobile phone assistant obtained authorization at the OS layer, it realized the automatic takeover of users' cross - application tasks, allowing the industry to see the prototype of a complete GUI Agent. The imaginative space brought by this form will not disappear with the "countermeasures" of some super apps.

On the other hand, with the obvious leap of large models in instruction understanding, multi - round planning, and tool - calling capabilities in the past two years, AI has for the first time had the possibility of undertaking complex task chains. At the same time, the continuous improvement of mobile computing power, memory, and local inference capabilities has also made "edge - side execution" no longer just a concept.

One of the keys to making AI "runnable" on mobile phones is the leap in the NPU peak computing power and energy efficiency of mobile SoCs in the past two generations. According to the analysis of the industry research institution Counterpoint, in 2025, almost all mainstream global flagship - level SoCs fully supported generative AI. About 88% of the high - end SoCs shipped this year already had generative AI capabilities, and the AI computing power peak of these chips was close to 100 TOPS in 2025, about four times the level in 2021.

Under the wave of AI "boarding" mobile phones, the logic of voice assistants mainly based on fixed instructions in the Siri era can no longer meet users' expectations for "intelligence".

In such a window period, mobile phone manufacturers urgently need to find new experience breakthroughs at the hardware and system levels, while AI cloud manufacturers have more mature model and engineering capabilities. The convergence of the two parties is not an accidental business exploration but the result of the superposition of technological maturity and industrial demand.

This is why, although the sales volume of the M153 is limited, the 30,000 engineering machines are still like a signal flare, giving a charge signal to the mobile phone industry that has long been eager to move.

In this context, it is reasonable for the market to spread the news that larger manufacturers like vivo are promoting cooperation with ByteDance. This does not mean that all manufacturers will copy the deep - customization model of the M153. Vivo previously released its "Blue Heart Intelligence" AI strategy and self - developed Blue Heart large - model matrix, and emphasized the upgrade of OriginOS and system - level intelligent agent capabilities.

On the other hand, taking Honor as an example, the YOYO Agent has accessed some capabilities of the Doubao large model, but the training, product logic, and system integration of the Agent are still led by Honor. This way of "accessing capabilities rather than outsourcing products" may be the realistic choice for more mobile phone manufacturers at this stage.

It is worth noting that Luo Wei, the chief imaging engineer of Honor, previously expressed his personal view on the Doubao mobile phone assistant: "Since ByteDance has no hardware R & D capabilities, or its hardware capabilities are not more advanced than ZTE's. Then two short people together won't give birth to a tall one."

However, after this evaluation, Honor's booth still appeared at this Prime Conference as a partner of Volcengine.

Looking back, ByteDance's radical move is not a reckless advance out of step with the industry rhythm but is based on a gradually mature window period. The significance of the M153 does not lie in proving the success of a certain product but in making the industry more firmly see that the edge - side Agent has become an unavoidable path.

Facing the increasingly high walls built by apps such as WeChat and Taobao, the relationship between apps and Agents has become the next unavoidable question.

2

Since the launch of the Doubao mobile phone assistant, it has quickly sparked some discussions in the industry: When AI can directly understand intentions and perform operations on behalf of users, is the Agent a threat to traditional apps?

During the group interview session of this Volcengine Prime Conference, Tan Dai responded to this question raised by Facing AI. He first elaborated from the user's perspective.

"If a user wants to order a cup of coffee, whether this operation is completed through voice interaction with a robot or by operating the mobile phone, the ultimate goal is to meet the user's needs." In his view, ordinary people still have many unmet needs, and AI can better solve these problems.

"As for whether the carrier of this AI is a Web, an app, an agent, or an agent in an app shell, it may not matter that much." Tan Dai said.

Therefore, when it comes to the native mobile phone assistant, the number of apps that can be operated may never be the core of the problem. What really matters is whether AI can release the previously unmet needs with a lower cost of use. When the needs are "expanded", it will undoubtedly establish a new user mentality at the hardware level, especially in 2025 when the majority of users are no longer unfamiliar with AI.

QuestMobile data shows that as of the first quarter of 2025, the monthly active user scale of native AI applications has expanded significantly. Users' attitude towards native AI applications has gradually changed from "trying out" to "relying on", and the year - on - year growth rate of the average monthly usage times per user is significantly higher than that of the usage duration.

The report shows that the average monthly usage times per user of native AI applications increased by about 53.1% year - on - year, while the average monthly usage duration per user increased by about 32.7%. Users tend to open AI applications multiple times for short periods to meet instant needs.

These needs are exactly what the native mobile phone Agent can amplify. But this is essentially an incremental logic, which does not mean that the value of apps will be rapidly weakened in the short term.

The Digital Application Ecosystem Report released by the China Academy of Information and Communications Technology in 2024 pointed out that users' usage time is still highly concentrated on a few top - tier apps. Core applications such as payment, instant messaging, and short - video together account for more than 80% of the total user usage time. The account systems, payment capabilities, content supply, and security responsibilities carried by these apps do not have the conditions to be directly replaced by Agents in the short term.

That is to say, the Agent is changing the "entry position" of apps, rather than eliminating their existence value.

In this structure, apps are still encapsulations of capabilities and services, but they are no longer necessarily the starting point for users to express their needs. The Agent gradually undertakes the role of the "demand scheduling layer", responsible for understanding intentions, disassembling tasks, and distributing executions to different applications or system capabilities.

Moreover, the terminal through which users drive the Agent may not necessarily be a mobile phone.

According to the tracking and analysis data of Counterpoint Research, in the smart wearable device market in 2025, terminals with local AI capabilities, such as AI smart glasses, showed a significant growth trend. In the first half of 2025, the global shipment of smart glasses increased by more than 110% year - on - year, and the shipment proportion of AI - type smart glasses increased significantly, reflecting the industry's rapid adoption trend of local AI capabilities.

In this context, ByteDance will not be the only large - model manufacturer to enter the market.

Zhipu AI recently open - sourced its core AI Agent model, AutoGLM, enabling researchers and developers to build AI terminal intelligent agents with "Phone Use" capabilities based on this model.

AutoGLM can understand screen content and complete multi - step task processes by simulating user operations (such as clicking, swiping, and text input). It has currently demonstrated the ability of automated operations in more than 50 high - frequency Chinese application scenarios, such as WeChat, Taobao, Douyin, and Meituan.

In other words, when more large - model manufacturers try to "board" mobile phones, the dividing line is no longer just model access, but whether mobile phone manufacturers are willing to cede the AI entry and rules at the system level. The next - stage competition is more like a contest between "system and ecosystem".

3

If the nubia M153 is more like an "engineering verification", then the progress of Xiaomi's Pengpai OS and AI disclosed at the Xiaomi Smart Eco - Conference a few weeks after its release represents the synchronous follow - up of traditional mobile phone giants.

At this conference, Xiaomi did not focus on a single model or function but repeatedly emphasized the role of Pengpai OS as a system - level foundation. It is becoming a unified operating system framework connecting multiple terminals such as mobile phones, tablets, and TVs, and is also the core level where AI capabilities are organized, distributed, and restricted.

This judgment is not just at the conceptual level. According to the data disclosed by Xiaomi, as of the end of 2024, the number of connected devices on the Xiaomi AIoT platform exceeded 820 million, covering multiple categories such as mobile phones, tablets, TVs, wearables, and smart home devices. The goal of Pengpai OS is to integrate these scattered devices into a unified system framework, so that AI capabilities can be scheduled and reused among different terminals, rather than being attached to a single application.

This means that Xiaomi chooses to firmly place AI capabilities within the operating system, and the corresponding hardware ecosystem of this system is also huge.

On the other hand, Luo Fuli, the person in charge related to Xiaomi's MiMo large model, made her first public appearance at this conference. When talking about the evolution direction of AI capabilities, she focused on the task understanding and execution level, emphasizing that AI needs to have the ability to disassemble, plan, and cooperate with system capabilities to complete complex needs, rather than just staying at the stage of single - round Q&A or information generation.

From the perspective of the industrial structure, this route forms a contrast with the path of AI cloud manufacturers represented by ByteDance.

The logic for leading AI manufacturers to enter the mobile phone Agent market is not complicated because mobile phones are still one of the most important and stable high - frequency entry points for users. However, for mobile phone manufacturers, the operating system, hardware integration, and ecosystem control are their long - term advantages. Once the Agent is deeply embedded in the system layer, who defines the interaction mode, permission boundary, and default entry determines the final dominance.

This is why, while most leading mobile phone manufacturers access external model capabilities, they still adhere to self - developing Agent frameworks and system - level assistants. They prefer to regard cloud manufacturers as capability providers rather than product definers.

In fact, this is not a newly emerged path. Earlier, when Huawei was promoting HarmonyOS and the Xiaoyi assistant, it continuously strengthened system - level scheduling and cross - device collaboration capabilities, and evolved AI capabilities as part of the operating system.

According to the public data released by Huawei at the 2024 Developer Conference, the HarmonyOS ecosystem has covered more than 800 million devices, with more than 2.2 million registered developers, and runs on mobile phones, tablets, wearables, in - vehicle systems, and various IoT terminals. In this system, Xiaoyi is not an independent application but deeply participates in system - level scheduling and cross - device collaboration to uniformly call the capabilities of multiple terminals.

In new hardware forms such as smart glasses and earphones, this ecological difference is further magnified. For devices with limited screens and fragmented operations, the Agent is closer to the "basic interaction layer", and brand and system collaboration capabilities are often more important than a single model capability.

In this regard, obviously traditional hardware manufacturers have greater advantages. However, AI manufacturers such as ByteDance and Alibaba are also promoting their respective AI hardware businesses. It is still difficult to determine whether Xiaoyi and Super Xiaoai can comprehensively surpass Doubao in terms of edge - side experience.

But the recent "increased investment" in Doubao and other AIs in the mobile phone circle clearly shows that both AI manufacturers and mobile phone manufacturers have set their sights on the still - untapped system - level space of AI mobile phones.

Driven by the "signal flare" of the Doubao mobile phone assistant, 2026 may be a year for AI mobile phones to "grow and strengthen".

This article is from the WeChat official account "Zimubang" (ID: wujicaijing), author: Li Zhaofeng, published by 36Kr with authorization.