REVIEW

A survey on edge multimodal large models: Compression, inference acceleration, and applications

College of Control Science and Engineering, Zhejiang University, Hangzhou 310027, China

^* Corresponding authors (emails: This email address is being protected from spambots. You need JavaScript enabled to view it. (Yuanchao Shu); This email address is being protected from spambots. You need JavaScript enabled to view it. (Jiming Chen))

Received: 31 January 2026
Revised: 19 April 2026
Accepted: 21 April 2026

Abstract

Deploying multimodal large language models (MLLMs) at the network edge is critical for enabling low-latency, privacy-preserving multimodal intelligence. However, the substantial computational and memory demands of MLLMs present significant challenges for deployment on heterogeneous and resource-constrained edge devices. This survey systematically reviews existing approaches aimed at addressing these challenges. We categorize the literature along two complementary dimensions: model-level compression, which focuses on efficient architectural design and parameter reduction, and system-level inference acceleration, which emphasizes runtime optimizations such as scheduling and resource management. In addition, the survey examines the practical applications of edge-deployed MLLMs in domains such as cyber intelligence and embodied intelligence, and discusses emerging research directions, including edge-native model architectures, to further improve the trade-off between intelligence capability and resource efficiency.