On September 16 this year, the Nubia NaviX Ultra, equipped with the consumer version of the Doubao Phone Assistant, was officially released and went on sale simultaneously.
On the first day of sales, official reports from Nubia showed that sales revenue exceeded 100 million CNY within one second of the available stock going on sale. The significance of this figure lies not in the "speed," but in marking a deeper turning point: AI agent smartphones have finally moved from the "exhibition stage" of engineering prototypes and technology previews to the "shelf stage" of scaled mass production and commercial delivery.
01. From Engineering Prototypes to Mass Production and Commercialization: Agent Smartphones Complete the "Last Mile"
Looking back at the development path of agent smartphones, 2025 was the year of "proof of concept." In December 2025, the technology preview version of the Doubao Phone Assistant was launched. The engineering prototype Nubia M153 equipped with this version went on sale at a price of 3,499 CNY, and 30,000 units were quickly sold out. The market feedback was enthusiastic, but it also clearly exposed the shortcomings of industrialization: between the prototype and mass production, the gap is not technology, but the maturity and yield rate of the entire supply chain.
This 2026 release crossed exactly this threshold. Released at 14:00 on September 16 and going on sale the same day, it was simultaneously rolled out across channels such as ZTE Mall, JD.com, Tmall, Douyin, and Pinduoduo, with first-day sales breaking 100 million CNY in one second. The new phone is available in three versions: 12GB+512GB, 16GB+512GB, and 16GB+1TB, priced at 5,999 CNY, 6,499 CNY, and 7,499 CNY respectively; among them, the 12GB+512GB version supports the national subsidy, with a final price starting from 5,499 CNY. With sufficient inventory, smooth channels, and stable supply, the product already meets all conditions for scaled delivery. According to Nubia, from the technology preview version to the mass-produced model, the team collected and organized feedback from over 2,000 users, continuously optimizing functions and user experience.
What is even more noteworthy is the realization of "commercialization." The biggest doubt about agent smartphones previously was that "the demos are stunning, but they are impractical for daily use." This time, the product is sold directly to ordinary consumers. The target users are no longer limited to digital enthusiasts and industry early adopters, but cover a broader population such as working professionals and students. This shift in user structure from "technology early adoption" to "daily tools" is the true hallmark of mass production and commercialization.
From an industry perspective, the value of this step even exceeds the product itself. The fact that a product can achieve scaled mass production and commercialization means that a whole set of chains, including chip adaptation, on-device model deployment, system-level permission integration, and application ecosystem coordination, has been successfully established. This is not a single-point breakthrough by a single enterprise, but the collective maturity of a complete industrial collaboration chain.
02. "Understands, Executes, Remembers, and Secures": Full Delivery Across Four Dimensions
If mass production is the "form," then capability is the "substance." According to Nubia's official introduction, the Nubia NaviX Ultra focuses on four dimensions: "understands, executes, remembers, and secures," which correspond to response, execution, memory, and security, respectively.
Let's start with "understands." The device integrates the Seed full-duplex voice large model, supporting continuous conversation with interruptions at any time. Once the continuous conversation mode is enabled, a single wake-up allows for multiple rounds of continuous commands. According to Nubia, in scenarios where the device is equipped with the Doubao Phone Assistant, the AI wake-up rate in noisy environments is 48% better than competitors, and the sentence accuracy rate is 21% better than competitors, with support for the recognition of over a dozen dialects, including Cantonese and Minnan.
"Executes" is the core watershed that distinguishes agents from traditional voice assistants. Traditional assistants can only "respond," whereas agents can "execute"—they can invoke across applications, chain multi-step operations, break down a user's single request into a series of specific actions, and complete them one by one.
According to the Doubao Phone Assistant team at the launch event, the consumer version possesses four core underlying capabilities: deep reasoning to understand complex instructions and autonomously decompose and plan execution paths; generalization capability to achieve cross-application knowledge transfer, drawing inferences from one instance to adapt to unfamiliar scenarios; proactive exploration for self-correction to flexibly handle uncertainties; and ultra-long context adherence to support long-chain operations that always stay aligned with the user's initial goals. Nubia stated that the end-to-end task success rate exceeds 80%. What does this number mean in real-world scenarios? It means that out of ten tasks assigned by a user, eight can be completely and correctly executed. This has already crossed the threshold from a "toy" to a "tool."
"Remembers" solves the long-standing "amnesia" problem plaguing assistants. The device supports multiple manual methods to save screen content, such as voice input, three-finger swipe up, or shortcut keys, forming a "personal memory bank" that can be queried and answered at any time; with user authorization, it can also remember preferences proactively shared by the user, making task execution more aligned with personal habits. Local data also supports natural language retrieval. With a simple command like "Help me check the pickup code for the express delivery on July 6th," the answer is instantly provided.
"Secures" is the prerequisite for the large-scale commercialization of agents. For agents to execute operations across applications, they must obtain system-level permissions; the greater the permissions, the heavier the security responsibility. The device is equipped with an independent AI button on the side of the body that integrates fingerprint recognition. According to Nubia, this is the industry's first independent button integrating fingerprint recognition with AI wake-up. Users can wake up the AI while pressing the button to complete fingerprint authentication, and the task is executed only after the user's identity is confirmed, balancing convenience and security. According to Nubia's official introduction, the device has obtained the filing for the agent large model and is equipped with a lock screen security strategy; data collection follows the principles of "reasonable necessity" and "user autonomous control," adopting a combined edge-cloud data processing solution and passing multiple international security certifications.
The four dimensions point to the same conclusion: the capability framework for agent smartphones has been established, and the next competition is no longer about "whether they have it," but "how user-friendly they are."
03. Ecosystem Co-construction: The SAEP Protocol and the Rule of "Apps Have the Right to Say No"
The real difficulty of agent smartphones has never been in the phone itself, but in its need to interact with a massive number of third-party applications. It is like a highly capable person who cannot get things done if they cannot enter others' doors or use others' tools.
Addressing this common industry challenge, the Doubao Phone Assistant introduced the Screen Automation Execution Protocol (SAEP) during the release of this consumer version. The protocol has been subject to a 30-day public notice of rules starting from September 14.
The core mechanism of this protocol is to hand over the decision-making power of "whether AI can operate this application" to the developers. According to the protocol, third-party applications can independently choose to allow or refuse the AI assistant to execute screen automation operations within the application, and can further restrict the scope of permissions—they can either prohibit all operations or set individual limits for specific operations such as screenshots, simulated input, and content modification. For example, a content community can allow the assistant to help users draft content, but the final "publish" action must be pressed by a human. This is the most essential difference between SAEP and the previous binary choice of "block or not block."
Regarding the execution standards during the public notice period, the rules are equally specific. During the notice period, the Doubao Phone Assistant only performs automated operations on built-in system applications, ByteDance applications, and third-party applications that have explicitly agreed to access via SAEP or email; all other applications are not operated by default. After a developer submits a refusal declaration, the relevant application will be added to the inoperable list.
The handling after the expiration of the public notice period also leaves room for flexibility. Applications that have not declared refusal through official public notice channels and have not configured disabling strategies will have operations gradually opened according to risk levels; whereas third-party applications that explicitly refuse will never be within the scope of operations. This design of "silence does not equal refusal, and refusal is the final outcome" provides a buffer for onlookers and certainty for those who refuse.
In view of this, ecosystem construction remains a slow process that requires extensive participation from developers. However, establishing boundaries at the rule level means that a replicable model has been found for this path. As more and more applications connect according to the same set of rules, the capability boundaries of agents will continue to expand, and the user experience will enter a positive cycle. This is the inevitable path for agent smartphones to move from "standalone intelligence" to "ecosystem intelligence."
04. Domestic Supply Chain Synergy: Industrial Resonance from Storage to Displays
The mass production of an agent smartphone involves the entire domestic supply chain. From chips to storage, from displays to structural components, and from sensors to acoustic devices, every link must meet the new requirements under agent scenarios.
The new hardware demands posed by agents are clear. Cross-application task execution brings higher computing power and memory consumption; on-device model deployment sets higher standards for storage read/write speeds and capacity; and long-duration, high-frequency interactions pose new tests for heat dissipation and battery life. Specifically for this device, according to Nubia's official introduction, it is equipped with the 5th Gen Snapdragon 8 Elite mobile platform deeply optimized for AI with Qualcomm, featuring a 6.78-inch display, a built-in 7100mAh battery, and supporting 90W wired fast charging and 50W wireless fast charging. Behind these specifications are targeted adjustments made by the entire supply chain around the new scenario of high-frequency AI invocation.
It is worth noting that in the critical link of storage, the device adopts LPDDR5X products from CXMT, Samsung, and others, with maximum speeds reaching 10,667 Mbps, providing ample computing power for multi-task parallel scheduling and cross-application linkage. High-bandwidth memory is no longer just icing on the cake for agent smartphones, but a hard prerequisite for continuous cross-application switching and unbroken long-chain tasks—one command given involves over a dozen application switches and page loads behind the scenes, and any single lag will cause the task to fail.
From a more macro perspective, the emergence of agent smartphones has opened a new growth channel for the industrial chain. It brings not only the sales volume of a single category but also the overall upward shift of the entire set of hardware specifications—larger memory, faster storage, stronger heat dissipation, and longer battery life. The driving effect of this "specification upward shift" on the industrial chain is far more profound than that of a single model itself.
Of course, we must also clearly see that the scaled commercialization of agent smartphones has just begun. The perfection of the application ecosystem, the cultivation of usage habits, and the integration of security rules still have a long way to go. But starting from the moment of "hitting 100 million CNY in one second," the industry has stood at a new starting point.
AI agent smartphones are no longer just a concept, but real, purchasable products on the shelves that can be used in daily life.