vivo Launches Four BlueLM Models, 30B On-Device Model in the Works

At the 2026 vivo Developer Conference, vivo launched four BlueLM models. The BlueLM-RealTime voice model uses an end-to-end speech understanding architecture, handling noisy environments, accents, and mixed Chinese-English speech. The BlueLM-Nano on-device model has 3 billion parameters and can understand screens, images, videos, and text while remembering user habits and preferences long-term. It ranked first in multiple authoritative evaluations for perception and memory. The BlueLM-Flash/Pro cloud models use a self-developed Agentic engine and system-level Harness, acting as a phone task controller that executes user intents across devices and apps.

vivo also introduced the system-level BlueLM Harness, which reaches deep into the original system kernel and integrates over 6,000 atomic skills supporting more than 10,000 task combinations. vivo built a system-level perception and personal memory library for each user, providing continuous and controllable personal context for agents.

vivo revealed it is pre-researching a BlueLM 30B MoE on-device model for phone applications.