According to an announcement made by the China Academy of Information and Communications Technology (CAICT) on their official WeChat account, a significant initiative has been launched: To address issues such as inconsistent service quality and difficulties in model selection within current MaaS (Model as a Service) platforms, CAICT officially introduced a series of MaaS standards.

As large AI models deepen their integration into various industries, adapting these foundational models through engineering development for specific fields and creating intelligent applications for end-users has become a critical challenge in the final phase of implementing large models. In recent years, while numerous MaaS platforms have emerged in the industry, several problems have surfaced, including inconsistent model service quality, platform selection difficulties, and complex customization needs. Therefore, there is an urgent need for a comprehensive set of standards to guide the healthy and orderly development of the MaaS sector.

Keeping pace with the latest trends in the MaaS domain, CAICT established the MaaS Working Group in 2023 under the auspices of the China Artificial Intelligence Industry Development Alliance, collaborating with experts from academia, research institutions, and enterprises to initiate the drafting of the MaaS standards. After over a year of dedicated efforts, this series of standards has now been officially released.

The standards target practical challenges encountered during the implementation of MaaS, proposing specific capability requirements across multiple dimensions such as customized optimization of large models, service deployment and inference acceleration, model management, and application development. These standards not only serve as crucial references for large model service providers and platform suppliers in building capabilities but also provide robust bases for technical selection by application parties.

This series comprises six parts:

1. Model Service Agreement Requirements (AIIA / PG 0110-2024): Defines key elements like model inference performance, service availability, accuracy of measurement, and terms of responsibility, offering important guidance for both parties when signing service agreements.
2. Model Service Evaluation Methods (AIIA / PG 0173-2024): Establishes evaluation criteria and methods for model service quality, providing scientific bases for assessing compliance with service level agreements (SLAs) and assisting application parties in making informed technical choices.
3. Model Platform (AIIA / PG 0134-2024): Standardizes full-stack capabilities ranging from data engineering to service operations, reducing technical barriers for customizing domain-specific models and enhancing overall efficiency and capability levels.
4. Model Service Platform (AIIA / PG 0137-2024): Specifies the entire process capabilities of large model service APIs, guiding platform suppliers in strengthening their capabilities to improve deployment efficiency and service quality.
5. Model Management Platform (AIIA / PG 0193-2024): Clarifies management and operational requirements for AI assets and promotes better utilization of large models through mechanisms like model cards, fostering synergy between large and small models.
6. Model Application Development Platform (AIIA / T 0196-2024): Standardizes the complete process of developing AI applications based on large models, from component construction to platform operation, helping enterprises swiftly create proprietary large model applications.

Official sources revealed that since the launch of the MaaS assessment series in the second half of 2024, preliminary evaluations of certain modules for seven companies—iFLYTEK, Inspur Cloud, Huawei Cloud, Whale Cloud, Yuntin, Unisoc, and China Unicom—have been successfully completed. New rounds of assessments are currently ongoing.

The release of this series of MaaS standards marks a significant step forward by CAICT in promoting the healthy development of the MaaS industry, providing substantial support for the future growth of the MaaS field.