# MagicVL-2B: Empowering Vision-Language Models on Mobile Devices with Lightweight Visual Encoders via Curriculum Learning

**Authors:** Yi Liu, Xiao Xu, Zeyu Xu, Meng Zhang, Yibo Li, Haoyu Chen, Junkang Zhang, Qiang Wang, Jifa Sun, Siling Lin, Shengxun Cheng, Lingshu Zhang, Kang Wang  
**arXiv:** [2508.01540](https://arxiv.org/abs/2508.01540)  
**Date:** 2025-08-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.01540v1/

## Key points

- Vision-Language Models (VLMs) have advanced in various aspects of everyday life
- Computational and storage requirements of VLMs pose challenges for mobile deployment
- MagicVL-2B introduced as a novel VLM optimized for flagship smartphones
- Features a lightweight visual encoder and redesigned dynamic resolution scheme
- Multimodal curriculum learning strategy proposed to enhance compact encoder performance
- MagicVL-2B achieves comparable accuracy to state-of-the-art models with 41.1% reduced power consumption
- Practical and robust solution for real-world mobile vision-language applications

## Summary

In recent years, Vision-Language Models (VLMs) have made significant advancements in various aspects of everyday life. However, their computational and storage requirements pose challenges for efficient deployment on mobile devices. To address this issue, a team of researchers led by Yi Liu introduced MagicVL-2B - a novel VLM optimized for flagship smartphones with a lightweight visual encoder and redesigned dynamic resolution scheme. They also proposed a multimodal curriculum learning strategy to enhance the performance of the compact encoder within VLMs. Evaluations on standard benchmarks showed that MagicVL-2B achieves comparable accuracy to state-of-the-art models while reducing power consumption by 41.1%. This makes it a practical and robust solution for real-world mobile vision-language applications, opening up new possibilities for enhancing user experiences and expanding the capabilities of mobile devices.