MaaS框架与应用研究报告（2024年）

Group 1 - The report outlines that Model as a Service (MaaS) is a new AI service model that packages AI models and related capabilities into reusable services, enabling enterprises to quickly and efficiently build, deploy, monitor, and invoke models without the need to develop and maintain underlying infrastructure [28][25][35] - MaaS is driven by the rapid development of large models, which exhibit significant performance improvements and broadened application ranges, but also face challenges related to high technical and economic costs [31][26][29] - The report emphasizes that MaaS can significantly lower the technical barriers for users, promote resource sharing in the industry, and facilitate the integration of model services into business scenarios [35][38][66] Group 2 - The current state of the MaaS industry is characterized by a preliminary formation of an industrial map, with various platforms providing end-to-end services from data processing to model training, validation, deployment, and monitoring [40][69] - Major cloud service providers like Alibaba Cloud, Tencent Cloud, and Baidu are actively developing platforms that support various machine learning algorithms and large models, offering low-code development environments and efficient model training and deployment capabilities [43][71] - The report identifies that the financial industry has become the leading sector for MaaS implementation, accounting for 49% of all industry applications, due to its established technological foundation and experience in traditional AI deployment [139][141] Group 3 - The report discusses the challenges faced by MaaS, including the lack of standardization in service quality and compliance management, which hinders the effective implementation of MaaS solutions [79][80][81] - It highlights the need for a robust compliance management system to address data privacy and responsibility issues, ensuring that data used for model training and optimization is legally compliant [80][81] - The report also notes that the ease of use of model services is currently insufficient, with a lack of transparency in model information and weak interpretability, making it difficult for users to select and understand the models [52][79] Group 4 - The MaaS framework consists of three layers: model platform layer, model service layer, and application development layer, each providing distinct capabilities that can be accessed independently by users [82][111] - The model platform layer offers a comprehensive toolchain for model customization and management, while the model service layer provides direct access to a variety of models, facilitating the democratization of large model usage [96][119] - The application development layer supports various integration methods for building AI applications, catering to users with different technical skills and enhancing the flexibility of application development [102][105] Group 5 - The report presents practical cases of MaaS implementation, such as the BankGPT platform developed by Ping An Bank, which allows users to efficiently utilize and develop models through API calls without needing to understand the underlying complexities [148][149] - It also discusses the ModelScope initiative, which encourages community contributions and provides a platform for model sharing, version management, and service hosting [155] - The report concludes that the rapid development of MaaS is enhancing the efficiency and effectiveness of AI applications across various industries, with significant improvements in model deployment and application customization [138][144]