Vivo Unveils End-Cloud Model Matrix and System-Level Harness for AI Phones

The company's starting point is a gap in AI phones: stronger model capabilities have not automatically produced better handset experiences. A flight delay notification, for example, can be summarized by AI, but rebooking, hotel changes and meeting adjustments require the phone to understand the user's current situation, remember past habits and coordinate models, context, apps and services under user authorization and privacy protection. vivo frames this as a system problem involving perception, memory, scheduling, execution and collaboration, not just a model benchmark problem.

At the AI session, vivo introduced the BlueLM end-cloud model matrix. BlueLM-Realtime, an end-to-end voice model, handles voice recognition and understanding of semantics, tone and intent. BlueLM-Nano, an on-device model, supports long-running perception and memory. BlueLM-Flash, a cloud-side model, executes high-frequency tasks such as information queries, calendar management and cross-app operations. BlueLM-Pro takes on complex reasoning and long-range planning. The system can also automatically call external leading models in specialized fields, turning models into background capabilities scheduled by the system rather than products users must select manually.

vivo cited Counterpoint Research's forecast that generative AI phones will account for 45 percent of shipments this year, up from 36 percent last year. It compared the end-cloud model arrangement to heterogeneous computing in chips: instead of using the strongest unit for every job, the system assigns tasks to the most suitable unit. The company said a single model cannot balance latency, cost, privacy and capability, so multiple models must work in concert.

For perception and memory, vivo said BlueLM-Nano uses a lightweight visual encoder and UI-specific tokens to understand screens, images, videos and text efficiently. For memory, it uses SwiftKV, hybrid sparse attention and PLE layer embeddings to extend on-device context to 32K. The company said perception is continuously deposited into memory, and memory makes the next service more tailored to the individual, creating a personal intelligence flywheel. It also stressed that continuous personal context must run within clear authorization boundaries.

On the engineering side, vivo introduced an Agent-native BlueLM system-level Harness developer platform. The platform has a perception and memory layer to understand users and environments, a planning layer for routing, task orchestration and reflection optimization, and an execution layer that connects models, agents, skills and external services to phones, IoT and the physical world through more than 6,000 system-level tools, as well as MCP, A2A, CLI and unified APIs. vivo said these layers form a closed loop of understanding, planning, calling and execution.

The Harness also includes a self-evolution mechanism. Based on an observe-reflect-precipitate framework, the system can continuously learn behavior and feedback under user authorization, review task paths during idle device time and feed new experience back into system capabilities. Unified intent, cross-device Runtime and context portability allow tasks to continue across phones, PCs and tablets, while MCP and A2A let agents, models and tools hand off to one another. vivo said personal AI accumulates not only static user information but also dynamic execution experience.

In the ecosystem layer, vivo positions itself between intent and services rather than trying to provide every service itself. After BlueLM XiaV was upgraded to Pro mode, vivo said developers can integrate once and have their capabilities understood, scheduled and reached through different system entrances. Agents, skills, MCP, atomic capabilities and on-device models are being brought into the same capability base. The competitive unit for intent services is shifting from a complete app to capabilities that can be intelligently scheduled.

vivo listed partners including Alipay, Meituan, Amap and JD. With Alipay, vivo said it restructured terminal service formats into service direct access, service execution and MCP Skill so more life services can be completed with one sentence. Meituan connected food, drink, entertainment and other capabilities to XiaV through an Agent. Amap is linked for navigation, route planning, life services and POI queries. JD connected cross-app selection, multimodal recognition, product search and payment. At the VDC roundtable, partners from JD and Alipay discussed real problems in landing Agent services on terminals and new solutions from ecosystem collaboration.