|
Work item:
|
F.CFM
|
|
Subject/title:
|
Requirements and framework for cloud-edge-end collaborative interaction between foundation models and lightweight models
|
|
Status:
|
Under study
|
|
Approval process:
|
AAP
|
|
Type of work item:
|
Recommendation
|
|
Version:
|
New
|
|
Equivalent number:
|
-
|
|
Timing:
|
2028 (Medium priority)
|
|
Liaison:
|
ITU-T SG13, SG17, SG20, and ISO/IEC JTC1/SC42
|
|
Supporting members:
|
Zhejiang Dahua Technology (China), China Telecom, Zhejiang University (China), Institute of Computing Technology Chinese Academy of Sciences, China Mobile, Zhejiang Lab (China)
|
|
Summary:
|
Foundation models are driving a paradigm shift in the field of artificial intelligence (AI)-based multimedia, demonstrating remarkable generalization capabilities in tasks such as natural language processing, multimodal understanding, and content generation across multimedia domains. These models are typically deployed in the cloud, leveraging powerful computational resources to provide general-purpose AI services. However, challenges such as high latency, difficulties in privacy protection, and limited adaptability to vertical scenarios remain to be addressed. In contrast, lightweight models are a type of AI models specifically designed and implemented for efficiency and low resource consumption. When deployed on edge or end-devices, they offer advantages such as low latency, high responsiveness, and enhanced adaptability to vertical contexts by effectively capturing local scene data. Nevertheless, their limited model capacity and generalization ability pose inherent constraints.
To integrate the broad knowledge of foundation models with the scene adaptability of lightweight models, cloud-edge-end collaborative interaction techniques have emerged. Through mechanisms such as knowledge distillation, capability transfer, joint optimization and inference, effective knowledge sharing and capability complementarity can be achieved between models of different scales. The goal of cloud-edge-end collaborative interaction is to combine the general intelligence of foundation models with the domain-specific strengths of lightweight models, enabling co-evolution through interactive learning and ultimately empowering downstream vertical AI multimedia industry applications.
This Recommendation specifies the requirements and a framework for cloud-edge-end collaborative interaction between foundation models and lightweight models, and provides related concepts, collaboration paradigms, use cases, and application scenarios. This Recommendation is intended to guide the design, development, implementation, and application of related systems or products.
|
|
Comment:
|
-
|
|
Reference(s):
|
|
|
Historic references:
|
|
Contact(s):
|
|
| ITU-T A.5 justification(s): |
|
|
|
|
First registration in the WP:
2025-09-09 11:24:30
|
|
Last update:
2026-05-08 11:32:57
|
|
