How OPPO’s AndesVL Is Revolutionizing On‑Device Multimodal AI

Technical Background

Current on‑device multimodal models suffer from limited performance, capability, and adaptability, hindering high‑performance, privacy‑focused, low‑latency AI phone applications. OPPO AI Center released AndesVL, an open‑source, full‑stack adapted multimodal model addressing these challenges.

Model Architecture

AndesVL adopts a unified architecture with four size tiers (0.6B‑4B) and supports both Instruct and Thinking modes. It combines a Vision Transformer encoder (AimV2‑300M or SigLIP‑2‑base), a multi‑layer perceptron, and a Qwen3‑based large language model, integrating 2D‑RoPE and NaViT for flexible resolution handling.

Training Scheme

The training consists of a pre‑training phase (visual‑language alignment, joint visual‑language training, and multi‑task training) followed by a post‑training phase (supervised fine‑tuning, mixed‑preference optimization, and reinforcement learning). Techniques such as MPO, GRPO, and extensive data filtering yield high‑quality multimodal data.

On‑Device Deployment Solutions

To enable efficient mobile deployment, OPPO applied model sparsification (up to 75% sparsity, 1.8‑bit BPW), quantization‑aware training (QAT) with a dual‑weight framework, and the QALFT method for LoRA‑specific quantization. Additionally, encoding compression (OKV) and speculative decoding (EAGLE‑2, HASS) achieve up to 6.7× decoding speedup and support 128K context length.

Evaluation Results

AndesVL attains top scores on over 30 benchmarks, surpassing competitors in overall ability, mathematical reasoning, visual‑text understanding, multi‑image comprehension, general QA, hallucination suppression, and multilingual tasks. Both 4B and smaller models demonstrate leading performance across diverse verticals.

Future Outlook

The team plans to advance visual encoder designs, post‑training strategies, knowledge distillation, and integrate vision‑language‑speech tri‑modal models, further pushing mobile AI capabilities.