vivo Outlines AI Phone Strategy With 30B On-Device MoE and System-Level Harness

After the event, Ifanr spoke with Zhou Wei, vivo vice president, vice president of OS and dean of vivo AI Global Research Institute, Guan Yanbing, general manager of vivo AI products, and Huang Zixun, AIOS product director.

Zhou said putting a 30-billion-parameter model on a phone would have seemed impossible a year earlier. At that time, vivo judged that running 3B or 7B models was already difficult and 13B nearly unfeasible because a 7B model could occupy more than 4GB of memory while flagship phones generally had around 12GB. This year, MoE changed the calculation. A 30B model can be split into many experts, and only a few are scheduled at a time; with loading and caching optimizations, two or three experts may require only 2GB to 4GB of dynamic memory, while inactive parameters can remain in 512GB or 1TB flash storage.

Zhou said the industry revised its judgment this year and vivo began training its own 30B MoE model. The model is still in engineering validation and is not yet ready for consumer phones. New bottlenecks include moving experts, chip scheduling, power control and on-device compute allocation. Zhou estimated the full capability would mature around 2028, when high-end phones are likely to stay at 12GB to 16GB of memory and only about 4GB can be permanently allocated to an on-device model.

Another term vivo emphasized was Harness. If a large model is the brain, Harness is the phone's execution hub: it maintains the main loop, finds tools, retrieves context, executes capabilities, verifies results and advances the next step. In complex real-world scenarios, BlueLM Xiao V has its own main loop and Harness has its own, and tasks could develop nested main loops or multiple Harnesses managed above. vivo embedded the Harness main loop into BlueLM Xiao V and further into system processes. Zhou said Harness itself is not model-specific; the underlying model can be DeepSeek, Claude or another and can be replaced by task. What should remain is the execution logic, context and user memory. For phone migration, he said the model can simply be downloaded again, while what needs seamless transfer is the perceptual memory, long-term habits and personal context accumulated on the old phone.

Guan summarized vivo's differentiation in personal AI in two points: an on-device Harness that can deeply schedule the entire OS, and a phone that holds user-specific context. Phones are among the few computing devices that accompany users 24 hours a day and accumulate photos, calendars, locations and device status. Rather than uploading all data to the cloud only after a user asks a question, vivo stressed pre-understanding on the device. Zhou gave a travel-planning example: a general AI assistant can plan a trip if a user provides enough details about an elderly father, a noisy child and a pregnant wife, but a phone that has long-term perception of family relationships and health conditions could automatically slow the itinerary and remind the user to bring a folding stool for the pregnant wife even if the user does not mention it.

To let Harness act, vivo needs tools. The company announced more than 6,000 atomic capabilities that can be flexibly orchestrated into more than 10,000 tasks. An atomic capability is the smallest independently callable function unit inside an app, such as note writing, format conversion, calendar creation or changing system settings. Huang said 70% to 80% of the 6,000 capabilities come from vivo's official system apps. In the past, users had to remember an app, open a page and press a button; for an Agent, vivo wants a path closer to Agent to Tool to function. The company is also packaging high-frequency combinations into Skills.

Third-party services such as Meituan, Alipay and mobile office tools are the larger battleground. vivo said it would not insist on a single protocol, citing MCP, A2A and CLI as different preferences. Huang said vivo sets no limit on access format; the core standard is what works more efficiently and friendly for an Agent. At the conference, vivo opened Agent, Skill, MCP and atomic skills, aiming to build a system interface layer that breaks app silos. Zhou said 6,000 capabilities are only a start; supporting the many Harnesses now on PCs may require 20,000 to 30,000 Skills and interfaces on phones, and the Blue River platform is moving toward that scale.

On GUI Agents, Huang said vivo is relatively conservative internally. GUI Agents can use visual recognition and simulated clicks when third parties do not open APIs, but Huang cited low efficiency, breakage after app redesign, cascading errors from small recognition mistakes and amplified security risks in payment authorization and private data. He argued that some app interfaces will become minimal or shift to pure background information services, which would reduce the long-term value of training AI to memorize which button to click. Zhou compared the issue to the PC ecosystem, where UI support for many private applications beyond the browser is weak because it is unnecessary. In the mobile ecosystem, GUI will remain for humans, while app capabilities can be provided to Agents through machine interfaces. He described a calendar example: instead of telling AI where the add or delete buttons are, the calendar module itself could sense an event and create an entry when news, messages or other services generate a clear scheduling need. vivo said it prefers apps to become Agent Ready. Guan added that MCP, Skills, Agents, A2A and GUI may coexist for some time, and GUI could be a fallback for long-tail apps such as government and life services in Alipay, though vivo prioritizes head services where most user pain points are.

vivo also extended the personal AI idea to AI office work with Xiao V Co-work. Zhou said vivo's positioning is not to compete on heavy coding or document production on a phone but to make the phone a control panel for a PC. Guan said that when people are away from a PC, such as in an interview room or traveling, they still carry a phone and may need to check how an Agent task is executing. Xiao V Co-work is therefore positioned as a control end that can remotely operate a work or home PC. The company ties the strategy to personal context on the phone, which it sees as a mobile-specific barrier.

Editor's Summary

vivo is positioning personal context, a system-level Harness and on-device models as the core of its AI Phone strategy. Its 30B MoE model remains in engineering validation and may not mature until around 2028, while the company is already opening thousands of atomic capabilities and multiple agent protocols. vivo is also taking a cautious approach to GUI agents and pitching Xiao V Co-work as a phone-based control panel for PCs.