Efficient Deployment of Multimodal LargeModels: A Surveyon Technical Innovations, Industrial Applications, and Challenges of Heterogeneous MoE Architecture, Low-bit Quantization, and Cloud-Edge-End Collaboration (2024-2026)

This survey focuses on the efficient deployment of multimodal large models (MLLMs) and systematically reviews the core technologies from 2024 to 2026. At the architectural level, heterogeneous Mixture-of-Experts (MoE) architectures represented by ERNIE 4.5 achieve a balance between cross-modal parameter sharing and modality-specific computation through modality-isolated routing and router orthogonal loss, significantly improving computational efficiency. In terms of model compression and inference optimization, methods such as Layerwise Ultra-Low Bit Quantization (LUQ) and MoEQuant effectively address the quantization challenges of multimodal data, while TensorRT-LLM and ONNX Runtime provide deeply optimized inference engine solutions for hardware architectures like Blackwell. Regarding deployment strategies, cloud-edge-end collaborative architectures represented by DistMLLM, AIVD, and NANOMIND meet the requirements of low latency and high privacy while achieving cost savings and energy efficiency improvements through adaptive offloading, dynamic resource scheduling, and hardware-software co-design. Application cases show that these technologies have effectively solved industry pain points such as high latency and data privacy in medical edge diagnostic devices and quality inspection \& predictive maintenance scenarios in smart manufacturing, balancing efficiency and performance.