Standard No.:	GB/T 45288.1-2025
English Name:	Artificial intelligence―Large-scale model―Part 1: General requirements
Chinese Name:	人工智能大模型 第1部分：通用要求
Professional Classification:	GB    National Standard
Source Content Issued by:	SAMR; SAC
Issued on:	2025-02-28
Implemented on:	2025-2-28
Status:	valid
Target Language:	English
File Format:	PDF
Word Count:	8500 words
Translation Price(USD):	255.0
Delivery:	via email in 1~3 business day

GB/T 45288.1-2025 Artificial intelligence―Large-scale model―Part 1: General requirements English, Anglais, Englisch, Inglés, えいご This is a draft translation for reference among interesting stakeholders. The finalized translation (passing through draft translation, self-check, revision and verification) will be delivered upon being ordered. ICS 35.240 CCS L 70 National Standard of the People's Republic of China GB/T 45288.1-2025 Artificial intelligence - Large-scale model - Part 1: General requirements 人工智能大模型 第1部分：通用要求 (English Translation) Issue date: 2025-02-28 Implementation date: 2025-02-28 Issued by the State Administration for Market Regulation the Standardization Administration of the People's Republic of China Contents Foreword Introduction 1 Scope 2 Normative references 3 Terms and definitions 4 Reference architecture 5 General requirements Bibliography Artificial intelligence - Large-scale model - Part 1: General requirements 1 Scope This document establishes the reference architecture for large-scale models and specifies the general requirements for large-scale models. This document is applicable to the development, preparation, deployment, and application of large-scale models. 2 Normative references The following documents contain requirements which, through reference in this text, constitute provisions of this document. For dated references, only the edition cited applies. For undated references, the latest edition of the referenced document (including any amendments) applies. GB/T 42018-2022 Information technology - Artificial intelligence - Platform computing resource specification GB/T 42755-2023 Artificial intelligence - Code of practice for data labeling of machine learning GB/T 45401.1-2025 Artificial intelligence - Scheduling and cooperation for computing devices - Part 1: Virtualization and scheduling 3 Terms and definitions The following terms and definitions apply to this document. 3.1 large-scale model large-scale deep learning model deep learning model trained on massive data, with complex computing architecture, capable of handling complex tasks and possessing certain generalization ability Note: The parameter count of a large-scale model is determined by its functionality and modality, generally not less than 100 million. The total amount of training data used for a large-scale model is influenced by its parameter count; the logarithm of the parameter count for a converged large-scale model is proportional to the logarithm of the total volume of its training data. 3.2 large-scale model service services for developing and applying large models and large model systems, as well as services providing support for the business activities of demand parties by means of these Note: Common large-scale model service content includes large-scale model platform services, large-scale model development and customization services, and large-scale model inference and operation services. 3.3 task scheduled training or inference object Note: A task is used to accomplish a relatively independent business function. A task belongs to and only belongs to one job. [Source: GB/T 25000.23-2019, 4.12, modified] 3.4 fine-tuning process of continuing training on a large model using domain-specific data to improve the prediction accuracy of machine learning models Note 1: Domain-specific data is typically production data or synthetic data from specific scenarios. Note 2: Common fine-tuning methods include prompt fine-tuning, full-parameter fine-tuning, parameter-efficient fine-tuning, etc. [Source: GB/T 41867-2022, 3.2.31, modified] 3.5 prompt prompt phrase instruction or information object inserted into input samples when using large models for fine-tuning or downstream task processing 4 Reference architecture The reference architecture for large-scale models from a functional perspective is shown in Figure 1, including the resource pool, tools, data resources, models, industry applications, and service platform/components. Among them: ——The resource pool includes hardware resources such as computing resources, storage resources, and network resources, as well as software resources such as resource virtualization/scheduling; ——Tools include data tools and model tools; ——Data resources include general-purpose data, domain-specific data, and private data; ——Models include foundational large-scale models and customized large-scale models. Foundational large-scale models can support unimodal or multimodal data. Customized large-scale models are large-scale models fine-tuned from foundational models according to user requirements to suit the production environment; ——Industry applications provide large-scale model downstream task matching services for users in various industry scenarios; ——The service platform/components run through all layers, providing support for the orchestration, deployment, model inference, operation, maintenance, and management of large-scale models and related services. 5 General requirements 5.1 Resource pool 5.1.1 Computing resources Physical computing facility components [such as Central Processing Unit (CPU), Graphics Processing Unit (GPU), Field-Programmable Gate Array (FPGA), Neural Processing Unit (NPU), Tensor Processing Unit (TPU)] or virtual computing devices that provide capabilities such as computation and data processing for model training and inference. a) Shall perform model training or inference for at least one modality (e.g., text, image, speech). b) Shall possess hardware acceleration functions for artificial intelligence computing, equipped with distributed training and inference computation acceleration libraries: 1) Training server: Shall have no fewer than 4 × 100GE network ports; power supply modules and fan modules shall support hot-swapping and redundancy (e.g., 2+2 redundancy, N+1 redundancy, etc.); 2) Inference server: Total memory bandwidth shall not be less than 800 GB/s; shall have no fewer than 2 Peripheral Component Interconnect Express (PCIe) expansion slots; power supply modules and fan modules shall support hot-swapping and redundancy (e.g., 1+1 redundancy, N+1 redundancy, etc.). c) Should possess hardware-accelerated preprocessing functions (e.g., image, video codec). d) Shall possess a key-value pair caching function.