vivo Announces Four BlueLM AI Models and System-Level BlueLM Harness Framework
At its developer conference, vivo unveiled four new BlueLM AI models, including the 3B-parameter BlueLM-Nano for on-device context and memory, BlueLM-RealTime for voice processing, and BlueLM-Flash/Pro for cloud tasks. Additionally, vivo introduced BlueLM Harness, a system-level framework integrated deep into the OS kernel that links over 6,000 atomic system skills for autonomous AI agent execution. This announcement highlights the mobile industry's shift toward embedding agentic AI directly into smartphone operating systems rather than relying purely on cloud chatbots. By providing AI agents with direct access to system kernel APIs and long-term user context, smartphones are evolving into proactive personal assistants capable of executing complex, multi-step workflows. BlueLM-RealTime utilizes an end-to-end Mixture-of-Experts (MoE) speech architecture to handle noisy, accented, or code-switched voice inputs, while BlueLM Harness pairs with an in-house Agentic engine to support over 10,000 task combinations. vivo also revealed that it is actively researching a 30-billion-parameter MoE model designed for on-device mobile execution.
## BACKGROUND
On-device AI relies on compact language models running directly on mobile hardware to reduce latency and enhance user data privacy. Agentic AI refers to systems capable of autonomously planning, breaking down goals, and executing actions across applications using tool calls rather than simply responding to text prompts.