Vivo Executives Detail 'Personalized Intelligence'; On-Device 30B MoE Expected Around 2028

After Vivo's developer conference in September 2026, vice president Zhou Wei, AI product general manager Guan Yanbing, and AIOS product director Huang Zixun discussed the company's 'Personalized Intelligence' strategy in an interview with ifanr. The strategy aims to embed AI into the system layer rather than delivering isolated features.

Zhou said Vivo is developing an on-device 30B Mixture-of-Experts model, which is still in engineering validation and expected to mature around 2028. In 2025, the team considered running a 13B model on a phone nearly infeasible, but MoE changed that view. With selective expert activation, dynamic memory usage can be kept between 2GB and 4GB, while inactive parameters reside in large flash storage.

Guan highlighted two differentiators: the on-device Harness can deeply coordinate the entire OS, and the phone holds a user's personal context. Vivo wants the on-device model to understand photos, files and device status in advance, avoiding the need to upload data to the cloud after each request, which improves efficiency and addresses privacy concerns.

Huang explained that Vivo has decomposed over 6,000 atomic capabilities that agents can orchestrate into thousands of tasks, with plans to expand skills and interfaces. On GUI agents, the company is currently cautious, preferring to make apps 'Agent Ready' through MCP, Skill and A2A protocols.

Zhou added that OriginOS remains the main carrier of Vivo's phone ecosystem, while BlueOS explores a more native agent environment, though its ecosystem is still maturing.