vivo held its 2026 developer conference in Shenzhen on September 16, releasing its BlueHeart intelligence strategy, OriginOS 7 and BlueOS 4, and compressing the whole story into one phrase: personal intelligence. Shi Yujian, senior vice president and chief technology officer, defined it as "understanding you, remembering you, getting things done."
The substance of the event is a shift in what phone AI is for. vivo's stated technical stack is "large models + Harness + agent security architecture": a three-layer model matrix — BlueLM-RealTime for speech understanding in noisy environments, accents and mixed Chinese-English speech; BlueLM-Nano for on-device perception and personal memory; and cloud models for complex cross-app, cross-device tasks — topped by the system-level BlueHeart Harness, which turns more than 6,000 atomic skills into tools agents can call, supporting over 10,000 kinds of tasks.
The most visible changes are system-wide AI features. The new BlueHeart Assistant Pro can decompose a task, integrate information across apps and operate across devices: one sentence produces a travel plan or a cross-device report. Screenshot recognition pushes tickets into the wallet and reminds of expiry; photo cleanup runs in a feed-like interface that learns preferences; file management aggregates, renames and summarizes; the keyboard recommends search terms and files from chat context. vivo's framing: "let important information appear proactively, let complex matters be completed by the system."
More notable is what vivo teased. Zhou Wei, vice president and head of vivo's AI global research institute, confirmed in post-event interviews that a 30B MoE on-device model is in deep pre-research. The bet rests on sparsity: with mixture-of-experts, the model does not need all parameters loaded at once; activating 1-3 expert sub-networks per request, 3-4GB of a flagship's 12GB memory could sustain dynamic operation. vivo's target is full local inference of a 30B-class model on ordinary flagships by the end of 2028, with no cloud involvement. For context, PC deployments of 30B models typically need 18-19GB of memory. If a phone can run one in 2-4GB, that implies instruction-level co-optimization between device makers and chip vendors.
vivo also made a deliberate route choice: it will not adopt GUI automation — simulated screen tapping — as its mainstream agent interface. Huang Zixun, director of vivo's AIOS products, argued that GUI operation carries high error rates and permission risks across many private apps. vivo has opened more than 6,000 system-level atomic capabilities and aims for 30,000. Guan Yanbing, general manager of vivo AI products, said basic AI features stay free for the general public, with data-package expansion only for heavy users.
OriginOS 7 and BlueOS 4 form the technical base. OriginOS 7, themed "comfortable intelligence," applies liquid motion effects across more than 200 high-frequency scenes and claims a 72-month anti-aging model; the first upgrade wave starts October 9. BlueOS 4 reports a 34% faster average memory allocation and 25% faster reclamation, and a full-stack Rust implementation across kernel, framework and apps; its AgentOS technical preview debuts on the vivo WATCH 6. On security, vivo assigns every agent a unique ID, minimum permissions, on-device sandboxing and private cloud computing, and says it holds the industry's first certificate for agent personal-information and user-rights control issued by CAICT, China's telecom research institute.
That a phone maker puts a 30B on-device model, a Harness and agent security into one strategy suggests on-device competition has moved past parameter counts. vivo's bet is that the next generation of phones will be defined not by hardware but by whether AI can finish tasks locally, privately and controllably. Whether that holds up is a question for 2028.
[1][2]