Are mobile AI assistants beginning to impose fees?

On September 16, vivo’s annual Developer Conference took place in Shenzhen. Alongside updates in areas such as the Blue Heart Intelligent Strategy, OriginOS 7, and BlueOS 4, it was announced that vivo’s AI assistant, Blue Heart Xiaov, will introduce a distinction between a standard mode and a Pro mode for task processing. Once the free quota for the Pro mode is depleted, users will need to purchase a membership or a Token Plan package.

These two modes serve distinct purposes. The standard mode of Blue Heart Xiaov primarily manages everyday high-frequency tasks, such as information inquiries, cross-APP operations, and schedule management. Conversely, the Pro mode is tailored for in-depth research, professional content creation, and complex long-term tasks. The mobile system then employs model routing to intelligently select and combine models based on task complexity, response time, call costs, and capability requirements.

Currently, mobile AI assistants are typically free for users but are limited to handling simple tasks. However, as Agent capabilities significantly enhance the handling of long-term tasks, AI provides a superior user experience but also incurs substantial costs, compelling smartphone manufacturers to explore more sustainable business models.

01 Why Are Smartphone Manufacturers Charging? Blame Agent for High Token Consumption

"Agent tasks consume an excessive number of tokens," stated Chui Wenwei, vivo’s AI Strategy and Technical Planning Director, to Shuzhi Qianxian. "Tokens have surged nearly 100-fold." The primary issue is that while the industry has achieved Agent capabilities, insufficient optimization has hindered further cost reductions.

Chui disclosed that query judgment and model selection, which are managed by vivo, do not consume significant resources. However, the current Agent framework, to ensure higher accuracy, keeps the Agent in a continuous loop, resulting in very lengthy AI contexts.

"Previously, it was a simple question-and-answer format, perhaps 10K, but now it can escalate to 100K," Chui explained. The loop volume and context volume follow an M×N process, consuming a substantial number of tokens. "If you ask 'hello,' your codex might consume 100K because it has to go through a loop regardless of the question."

vivo’s strategy is to categorize tasks. "After a query is received, if it’s a mobile-level request, I handle it. If it’s a general high-demand request, I find a suitable model for you," Chui said. For instance, vivo manages intent recognition and model routing, while simpler tasks like mobile function settings and multi-APP calls are also serviced by vivo’s models. However, complex reasoning and long-term tasks are delegated to third-party cloud-side models.

"We are continuously reducing costs throughout the entire chain," Chui informed Shuzhi Qianxian. In AI, effectiveness is prioritized first, and cost reduction is a matter of time.

Additionally, Shuzhi Qianxian learned that Blue Heart Xiaov will soon introduce video AI functionality, enabling users to interact with AI in real-time by opening the camera to the real world. Such interactive scenarios pose a significant challenge in terms of token consumption.

Therefore, vivo’s Pro mode will offer users a certain amount of free quota, with additional usage requiring token payments. However, the specific charging plan was not disclosed at the Developer Conference.

But in Chui’s view, this revenue does not go to vivo but is priced based on the public input and output prices of third-party models.

In fact, vivo’s approach is not surprising and closely resembles the business model of general-purpose Agent assistants on the market—simple tasks utilize Flash models, consuming fewer tokens and mostly being free. However, complex long-term tasks often require calling more advanced or external third-party models, necessitating payment to model manufacturers through purchasing credits or token packages.

Shuzhi Qianxian also learned from OPPO that its newly released AI assistant, Xiao Bu Next, demonstrated extensive proactive responses to long-term task processing and services on September 17. Currently, Xiao Bu Next is in the beta testing phase and is free for users to try. However, an OPPO engineer at the scene told Shuzhi Qianxian that a token Plan package would likely be introduced after the official launch.

02 Behind vivo’s Personal Intelligence: Testing Models and Harness

In the AI era, the way smartphones connect with users has undergone the most direct transformation: users previously had to decide which APP to open before finding functionality, but now they hope to accomplish these tasks with a single sentence.

Chui believes that to achieve personal intelligence, three elements must be simultaneously in place: first, understanding users and forming a continuously updated personal context; second, completing tasks, which relies not on a single model inference but on a continuous task process; third, connecting the ecosystem, as AI services cannot be provided by smartphone manufacturers alone and require developers to encapsulate AI capabilities into Apps.

Agent’s task-handling capabilities rely on the combination of large models and Harness. In vivo’s design system, the Blue Heart large model matrix provides reasoning generation and multimodal capabilities, while the Blue Heart personal intelligence framework connects systems and scenarios, linking perception, memory, interaction, planning, execution, and security to form a complete task closed loop.

Chen Xiaoxin, Senior Director of vivo’s AI Algorithm Center, provided an example: when a user wants to obtain a travel plan in Blue Heart Xiaov, the large model understands the goal and plans the itinerary, while Harness combines the user’s schedule, travel preferences, and historical memory to call services like weather and maps, transforming a general itinerary plan into a truly personalized one suitable for the user. "The large model defines the upper limit of intelligence, while Harness transforms intelligence into productivity."

However, a single model cannot simultaneously meet the needs of capability, latency, cost, and privacy. vivo’s approach is to construct a clearly divided and mutually cooperative Blue Heart large model matrix. Among them, the voice large model handles natural interaction, the on-device large model handles real-time perception and personal memory, and the cloud-side large model supports both the rapid execution of high-frequency tasks and the reasoning and planning of complex tasks.

At the same time, vivo has also integrated voice, on-device, cloud-side, sign language, and industry large models into the Blue Heart Intelligent Open Platform, which handles unified access, calling, evaluation, and governance. Developers do not need to manage complex model selection and combination and can focus more on scenario and product innovation.

Based on this technical architecture, vivo’s focus is on optimizing on-device and cloud-side experiences.

For example, on the on-device model, emphasis is placed on the collaboration between the model, system, and chip. At the chip level, vivo collaborates with chip manufacturers like Qualcomm and MTK on joint design, optimizes low-bit quantization and runtime decoding at the model level, optimizes memory scheduling and resource control at the system level, and collaboratively promotes operator optimization and heterogeneous computing power scheduling at the chip level. Ultimately, the first-word speed of the on-device model has increased by 30%, the initial speed has tripled, and screen perception has been reduced to 900 milliseconds.

Currently, vivo’s Itinerary Assistant, Xiaov Circle Search, AI Smart Notifications, and Album Agent are all built on the multimodal understanding and personal memory of the on-device model. It is reported that vivo is also pre-researching the Blue Heart 30B MoE on-device large model to explore completing long-chain complex tasks on smartphones.

According to Shuzhi Qianxian, these capabilities rely more on the smartphone manufacturers themselves to build. On the one hand, smartphone manufacturers can cooperate with their system-level tools to specifically optimize these tasks; on the other hand, these tasks often involve users’ personal data and privacy, which smartphone manufacturers are unlikely to entrust to third parties.

On the cloud-side model, vivo has focused on optimizing three technologies.

First, it deeply integrates the model with the Harness framework of vivo’s operating system, incorporating on-device memory, personal context, smartphone knowledge, and tool systems into model training to enhance the model’s understanding of smartphone scenarios and stable calling of system capabilities.

Second, it adopts the MOPD multi-expert teacher distillation mechanism to efficiently migrate different expert capabilities such as ToolUse, GUI Agent, Coding, and STEM into a unified model.

Third, it conducts online reinforcement learning based on real-machine environments, allowing the model to execute in real task environments and continuously optimize planning and execution strategies based on feedback.

"The ultimate goal is to use a smaller smartphone-specific model to handle users’ high-frequency mature tasks faster, more stably, and at a lower cost," Chen said.