Author | Ma Guangyu, Editor | Li Yuchen
"As large model capabilities grow stronger, what is the key to competing in the second half of the AI race?" On September 16, shortly after the opening of the 2026 vivo Developer Conference, Zhou Wei, Vice President of vivo, Vice President of the OS Domain, and President of vivo AI Global Research Institute, raised a new question. For the industry, this question is not unfounded. Over the past two months, divergence regarding AI in the industry has gradually become public. On one side, companies like OpenAI are collectively appealing that large model capabilities are already in excess, urging against involution-style development; on the other side, the market is beginning to question whether the capital expenditures of tech giants can be sustained. The smartphone industry, meanwhile, is fighting on another battlefield. Starting from the Doubao phone at the end of last year, "AI agent smartphone" has become a standard pitch at every launch event. This year, more than one manufacturer even claims to be the "world's first," accompanied by doubts about feature homogenization.
Over the past year, vivo's overseas shipment scale has continued to break records, achieving exponential growth in several key markets, with global users approaching 600 million. With such a massive user base, vivo is one of the few smartphone manufacturers capable of answering this question. At the main venue, vivo unveiled OriginOS 7, the all-new BlueLM Xiao V, the BlueLM 30B MoE on-device large model, Xiao V Cowork, and a technical preview of the BlueOS Agent Operating System, providing vivo's own answer: personalized intelligence. In the words of Shi Yujian, Senior Vice President and CTO of vivo, when a smartphone can perceive context, understand intent, and take proactive action, the Smart Phone will evolve into the Agent Phone. This is the first time vivo has directly defined the smartphone as an Agent at a developer conference. The core logic of this combination is "Large Model + Harness + Agent Security Architecture."
The large model is the "brain," responsible for listening, remembering, and executing. The voice large model has been upgraded to BlueLM-RealTime, the on-device BlueLM-Nano handles real-time perception and the accumulation of personal memory, while the cloud large model takes on more complex tasks. The Harness acts as the "nervous system," linking perception, memory, planning, and execution into an assembly line. It dives deep into the OriginOS kernel, turning the smartphone's over 6,000 atomic capabilities into the hands and feet of AI. Over the past nine years, the company has rarely adopted an aggressive stance on AI, and this shift in direction now sparks greater expectations from the outside world.
After the launch event, Zhou Wei, along with Guan Yanbing, General Manager of vivo AI Products, and Huang Zixun, Director of vivo AIOS Products, organized a small-scale communication session. Topics ranged from the memory pressure of the 30B model to the route debate between GUI and MCP, whether AI is in excess, whether the smartphone is the best carrier for on-device AI, and the pricing model for Agent smartphones. The following is the dialogue between Zhou Wei, Guan Yanbing, Huang Zixun, and media outlets including Lei Feng Network, edited for clarity without altering the original meaning:
Q: Your proposal to pre-research a 30B on-device large model this year was quite unexpected. Last year, your view was that it wasn't necessarily to pursue the parameter size of on-device models. What prompted this shift? Did you consider the memory pressure?
Zhou Wei: Last year, we felt that running 3B or 7B models on smartphones was very taxing on memory. But early this year, the industry changed. Large models don't necessarily need to be loaded all at once; each time, only one or a few expert models need to be loaded. Now, the standard memory for flagship models is 12GB, making it realistic to use 3GB of memory to load an MoE model. We revised our judgment on the industry and technology this year and quickly developed our own MoE model.
Q: Deploying the full-blooded 30B locally on a PC requires about 18-19GB of memory, but vivo believes the minimum is only 2GB. Will long-term use affect the experience? Can mid-range models enjoy the personalized 30B on-device experience?
Zhou Wei: This is a question we've been asking ourselves over the past six months. Deploying on-device models on laptops often requires a starting point of 64GB, which is impossible on a smartphone. However, with the emergence of sparse models like MoE, which load a portion of experts on demand based on task difficulty and activate 1-3B parameters, dynamically it may only require 2-4GB of memory. Theoretically, this is achievable, so we trained our own 30B MoE model. Why didn't we bring it directly to the product today? Because it poses entirely new challenges to smartphone hardware. To give a similar example, why was Liquid Glass proposed this year and not in 2024? Because the hardware chip was just planned in 2024, and it takes 24 months for a chip to arrive, making this year just right. Proposing MoE this year also means it will take another 24 months for the corresponding chip planning to materialize. Coupled with the fact that we still need a year or two to fine-tune the model, by the end of 2028, realizing today's so-called large model capabilities on-device with MoE will definitely become a reality.
Q: On-device models have grown from 1B a few years ago to 30B today. With parameters getting larger, what new requirements does this place on smartphone hardware?
Zhou Wei: I need to clarify something. The most precious resources on a smartphone are computing power and memory. By the end of 2028, flagship phones will have configurations of 12-16GB. The memory truly left for loading on-device models, I believe, is reasonably under 4GB. On-device models won't become massively large. The illusion that today's 30B is huge is just because it's sliced into 30 or 100 pieces, and not all of it is loaded when called.
Q: What exactly do the 6,000 atomic capabilities mentioned at the vivo Developer Conference refer to? Do they include MCP or multi-Agent?
Huang Zixun: Currently, the vast majority come from the system applications themselves. Atomic capabilities can be understood as how many functions an application provides, with each function directly atomized, such as writing notes or formatting and converting notes. Why are we determined to build atomic capabilities? Because in the future, complex user tasks will be entirely handled by AI for intent understanding and task orchestration. The orchestration process involves recombining subsets of local functions. The third-party part will be more flexible. Some leading office applications prefer using MCP, while some leading office Agents prefer COI or A2A. We don't have strict restrictions on formats; we do whatever is more friendly to the Agent's workflow.
Q: Have you considered opening up GUI operational capabilities? Last year, Mr. Zhou was still discussing privacy limitations, but this year many manufacturers support both MCP and GUI routes simultaneously.
Huang Zixun: Internally, we believe that in the long run, it will be weakly interface-oriented and weakly interaction-oriented; current GUI technology is very short-term. Even without GUI, we still have other more efficient technical means, which is also part of security. During past explorations, we also communicated with leading internet companies to hear their acceptance of the GUI approach. Therefore, today in our choices, we have weakened the GUI aspects significantly. Zhou Wei: When people use Harness on PCs, they will find that even on PCs, besides browsers, many private applications have very weak support for GUI. It's similar on smartphones; there's no need to be so obsessed with providing GUI support for third-party applications. We open standard PC Harness dependencies, supporting nearly 30,000 skills and interfaces, which is the scale already achieved on BlueOS. In the future, to run various Harnesses perfectly on-device, it basically needs to support this magnitude. But even after achieving this, it doesn't mean the smartphone must support GUI perfectly. Before December last year, to make the calendar support GUI, we had to add an "add calendar" function and a "delete calendar" function. This year is different. The calendar module itself will perceive event information; there is a demand, and it generates a calendar to-do item on its own. After 2026, supporting GUI is the application's own business; if they are willing to support it, they support it. It's no longer a topic of debate.
Q: When the BlueLM large model was released three years ago, it was mentioned that the cloud could reach up to 175 billion parameters. This year, the cloud focus has shifted to Harness and engines, and parameter size isn't emphasized much. Does this mean smartphone manufacturers believe the value of the cloud is no longer in the model itself, but in scheduling and execution?
Zhou Wei: Indeed, the perception of large models in 2025 and 2024 is completely different, and the language system of the entire ecosystem is different. In 2024, 70B was the mainstream cloud model, and 175B was the top-tier model, representing the industry ceiling. But from May to August this year, everyone's judgment on models went back and forth several times. Why do we want to make an MoE version of the on-device model? Because with the evolution of distillation technology, even a 3B model might be able to do a lot. The 3B from a year ago, this year, and three years later are not the same thing. Looking at it developmentally, the 3B in three years will catch up with today's 70B in many aspects. Today's 30B, in two years, might have each expert model equivalent to last year's 70B capability. This is the state of industry development.
Q: This morning, you showcased highly pain-point-solving features like Album Quick Clean and Ask & Answer. How does system-level AI innovation integrate with vivo's four long-term technology tracks? How can AI be made different from other manufacturers?
Huang Zixun: Internally, we have started to categorize which applications need to deeply integrate AI in future long-term development. The future trend is that interface elements will become fewer, interaction methods simpler, capabilities stronger, and even the need for interfaces will be increasingly weakened. As the system's Agent becomes stronger, some applications will gradually fade from view. What was showcased today are scenarios that will still exist in the long term but have massive pain points; we attempt to solve them through AI integration. For example, cleaning the album. Before the user triggers the function, the local on-device model has already understood the content. AI will make judgments based on the scenario, leaving the final step for user confirmation.
Q: Where exactly is the difference?
Huang Zixun: We are relatively determined in the industry to use on-device models for this kind of processing. In the past, users triggered functions, data went to the cloud, AI processed it, and then returned the results. Now, AI walks ahead of humans. When users see a scenario, AI has already done what needs to be done, leaving only confirmation. If this were done with a cloud model, it would mean that before the user even starts the function, all data goes to the cloud, leading to more privacy issues and not necessarily better results. Zhou Wei: At the AI level, vivo is the first in the industry to use on-device models on smartphones. Last year, we already used 18 modules, including the album. Now everyone's albums start at 100GB; running them on a server would take a long time. Why can the phone instantly complete "find a photo of taking kids to the beach during National Day"? Because the on-device model has already silently perceived these photos in the background and remembered the features. We have created a personalized storage space for everyone in the background. When you call it in the album, the support is already there. This wasn't possible before because photos had to be sent to the cloud for recognition. Now the on-device model does it silently in the background, ensuring privacy and security.
Q: Personalized intelligence emphasizes proactive understanding and proactive execution, but the boundary between proactivity and disturbance is very blurred and subjective. How does vivo define the proactive boundary of system-level AI? Are there measurement standards?
Guan Yanbing: This is an industry-level challenge. First, every user's demand scenarios and scale of good service are different, which is why we propose personalized intelligence. Only when the proactive service content and strategies differ for different users can this truly be resolved. Second, at the current stage, we will be more restrained in providing overly proactive services. For example, giving you a daily schedule or to-do suggestion, we will be more restrained in sending it initially. As users' positive behaviors like reading, saving, and importing to the calendar gradually increase, we will increase the depth and breadth of proactive services based on personal behavioral habits. Behind this is a set of experience standards evaluating whether user behavior is positive or negative. We are still exploring this issue.
Q: Recently, there have been two divergences regarding AI. One is that companies like OpenAI are appealing that large model capabilities are already in excess, urging against involution-style development; the other is questioning whether the capital expenditures of tech giants can be sustained. In your view, are today's large models sufficient or in excess?
Zhou Wei: AI is divided into three layers. The first layer is upward towards Artificial General Intelligence, where giants are currently competing. Moreover, I believe there is one piece of work that hasn't even started yet: self-evolution, self-iteration, and self-training. Everyone is just beginning this part. The claim that large model training has ended is invalid; a new journey has just begun. Second, today's large models can be used for productivity. Since July this year, most productivity scenarios have been sufficient. What truly limits you is imagination; the Harness cannot keep up with the wings of dreams, so the entire industry is working hard on this middle section. Third, the mobile end is a very important part of the digital world. For a long time, people still use PCs for productivity monetization, especially in 2B scenarios. But the smartphone is a device that stays close to us. Detached from the PC, you still need to check if the Agent's tasks and production are normal, which is why we built Cowork.
Q: Internet companies are also focusing on the Xiao V Cowork direction. As a smartphone manufacturer, how does vivo cooperate with them?
Guan Yanbing: Over the past two years, the most popular, widely applied, and deeply commercialized scenario for AI has been productivity and office settings. But our positioning is different. If your productivity is on a PC, when you leave your desk or travel, you always carry your phone. Our product positioning is to remotely control the Agent on your desk or home PC from your phone; it's a remote control panel. Based on this positioning, it will develop in a more portable direction in the future. Everyone has a pain point when using various Agents: having to repeatedly input background information and personal introductions. And the place with the most complete personal context is the smartphone. In the future, Cowork will provide richer personal context based on the phone's local Harness, allowing you to better use the Agent on your PC. This is our differentiation.
Q: What is missing for on-device AI to truly explode? Is it large models, communication computing power, or application explosion? Is the smartphone the best carrier for on-device AI?
Zhou Wei: We are essentially discussing whether capabilities should move down to the smartphone after the on-device shift of PCs and the cloud shift of servers. The industry trend we see today is that it will definitely come down. As for how many scenarios it can cover, it depends on industry development, which is just too fast. What I can see today is that MoE is a very good solution. Custom chips are currently very expensive, but computing power is not an issue. There are some people willing to pay a high price for a portable pure on-device computing power center and productivity center, which is fine, but this demographic and market are relatively small. This is a temporary view.
Q: AI functions have increasingly high requirements for computing power and cost, and some manufacturers have started using subscription models. Has vivo thought about the pricing model for Agent smartphones?
Guan Yanbing: Our current thinking is that the products we release will still provide free functions to users. But if some users want to use it very deeply and with very large traffic, we will provide a mechanism for extra payment to expand the traffic package, keeping our exploration going. I believe that in the coming months, the entire industry will have similar mechanisms. But for ordinary mass-market users, we will still provide a fairly usable free space.
Q: In 2024, you proposed the smartphone agent. Why didn't this thing move towards the C-end on a large scale in 2025?
Zhou Wei: The smartphone agent hasn't failed to move towards the C-end; rather, it has been internalized into many scenarios. For example, people have been ordering coffee for three years, and the smartphone agent has still contributed. A year or two ago, this capability looked like a direction. As we worked on it, we found that first, it was more mature than before, and second, the industry was debating it. Our attitude is to shelve the debate and let the ecosystem grow and mature naturally. Huang Zixun: We might not have used this term much in the past two years, but the mindset hasn't changed. In the past, it was application-level single-point intelligence, triggered by user voice. Now, it is system-level personalized intelligence, internalizing capabilities directly into the underlying layer. The underlying layer understands in real-time, performing local content understanding and intent prediction in advance, and executing on behalf of the user in controllable scenarios with authorization. We also made trade-offs, distinguishing which tasks are done by AI and which must be left to the user. But the concept of the smartphone agent proposed a few years ago hasn't changed.
Q: Today vivo released its self-developed BlueLM 30B on-device large model. There are also many excellent open-source models in the industry. How does vivo view the choice between self-development and open-source?
Zhou Wei: Even for trillion-parameter models today, as the threshold for pre-training lowers, as long as there is computing power, the model can be built properly. What helps us the most is distillation, but distillation is just a method. What truly gives us confidence and makes us feel it's worth investing in is seeing user pain points in this direction. In the future, when using large models, people will definitely say "help me adjust the brightness" or "what to do when carsick," and it will turn on the anti-carsick function for you. This kind of capability must be trained by smartphone manufacturers based on massive self-understanding and data accumulation. Besides data, we choose the on-device route because of another huge advantage: we do GPU optimization at the instruction set level. Manufacturers currently doing model training can't do on-device well; either the power consumption is too high, or the memory requirements are too high. Only by terminal manufacturers cooperating with chip manufacturers can on-device be done well. I believe that within three to five years, only terminal manufacturers making their own models on-device in this direction will provide the best experience. This is the reason we persist in investing.
Q: This year, manufacturers are all talking about Harness. Yesterday, a manufacturer also said their system was the first to truly implement Agent Harness. What are vivo's features and differences?
Zhou Wei: This question is actually the same as why we do accessibility and why only vivo has brought imaging to today's position. In our philosophy, this is a question of user orientation and scenario orientation. If this is most important to the user, even if everyone else does it, we don't just aim to be better than others; we must ensure the user is most satisfied in the scenario. Product competitiveness is not about comparing with competitors, but about whether we truly understand user pain points and scenarios. Why do so many visually impaired users use vivo phones for independent travel, independent living, makeup, picking clothes, and attending concerts? The core is that our team believes that doing things for such groups means we must create products that users prefer, that are different, and that are much better. If everyone is doing Harness, why should vivo still do it? You can look at the difference between vivo's imaging and the industry, and you can probably guess where our lead will be in two or three years for AI smartphones. It's hard to articulate in a few sentences where we do it well. But after a certain amount of time, everyone will feel, "This is a phone that satisfies me."
Q: What do you imagine as the ultimate endgame of personalized intelligence?
Zhou Wei: We always say that making a product ultimately means making it for a specific person or scenario. Just as we continuously do for the accessibility community, blind and deaf people will use our phones as their primary means for travel and life. Similarly, when we work for office workers and travelers, the scenarios must be made satisfying. A professional itinerary assistant cannot create a truly satisfying travel itinerary for you because it doesn't know you, your family relationships, social relationships, and behavioral preferences. But we can know you very well, even better than you know yourself. When BlueLM Xiao V creates an itinerary for you that exceeds the expectations of all family members, from that moment on, you will realize that you don't just use it for itineraries; you will also let it help with all kinds of things within its capabilities in life and work. Our vision is that in the future, a smartphone with personal intelligence is your exclusive personal assistant. This is our vision.
The author of this article has long focused on smart hardware and terminals. Industry professionals interested in terminal channels and industry dynamics are welcome to add the author's WeChat (ByArsT) for communication.