On 14 August 2026, China’s National Standardization Administration (SAC) announced the publication of three national standardization guiding technical documents on industrial artificial intelligence (AI): GB/Z 195-2026 Artificial intelligence — Reference architecture of industrial agent, GB/Z 203-2026 Artificial intelligence — Technical requirements of industrial foundation model, and GB/Z 204-2026 Artificial intelligence — System architecture of industrial foundation model. The three documents establishes an initial blueprint for industrial AI standardization.

GB/Z 195-2026 Artificial Intelligence — Reference Architecture of Industrial Agent

This document defines a reference architecture for industrial agents built around three core elements: functional modules, model resources, and interaction objects. The standard specifies eight capability modules—perception, memory, reasoning, planning, decision-making, execution, communication, and evolution—that enable an agent to perceive industrial environments, process information, formulate and execute tasks, communicate with other entities, and adapt to changing scenarios.

To support these capabilities, the architecture provides five model types accessed through dedicated interfaces: three large models (industrial base, domain, and scenario foundation models) that deliver progressively specialized capabilities from general industrial functions to industry- and scenario-specific tasks, and two small models (industrial AI small models and industrial mechanism models) that handle lightweight edge processing and physics-based deterministic simulation.

The agents are designed to interact with both the industrial physical space—people, machines, materials, production environments, and other embodied intelligent agents—and industrial information systems, including software, databases, operating systems, knowledge bases, and other non-embodied intelligent agents.

GB/Z 204-2026 Artificial intelligence – System architecture of industrial foundation model

This document establishes a system architecture for industrial foundation models organized into five layers plus security protection.

• The infrastructure layer covers cloud, edge, and terminal computing, storage, and communication resources.
• The data layer comprises industrial data and knowledge resources.
• The key technology layer addresses pre-training, adaptation, and inference deployment.
• The model layer defines a three-tier hierarchy of base, domain, and scenario foundation models.
• The application layer spans the full product lifecycle, including R&D design, production, testing, business management, and operations and maintenance services.

Security is treated as a cross-cutting requirement, with GB/Z 204-2026 mandating safeguards across data security, model security, deployment security, and service security throughout the model lifecycle.

GB/Z 203-2026 Artificial intelligence – Technical requirements of industrial foundation model

This document complements the GB/Z 2023 architecture with concrete technical requirements covering pre-training, adaptation and fine-tuning, inference and interaction, and deployment.

It sets capability benchmarks for industrial self-supervised and multimodal pre-training, cross-modal alignment, parameter-efficient fine-tuning, and mechanism-knowledge embedding. It further outlines requirements for retrieval-augmented generation, collaborative reasoning between large and small models, model compression, and hardware-accelerated inference. Deployment requirements distinguish between private cloud or local setups and public cloud platforms, specifying metrics for network latency, storage bandwidth, disaster recovery, and data encryption. These requirements are intended to ensure that industrial foundation models can be developed, adapted, and deployed in a consistent and verifiable manner.

GB/Z 204-2026 and GB/Z 203-2026 function as complementary documents. GB/Z 204 provides the architectural blueprint, defining what an industrial foundation model system looks like and how its components are structured. GB/Z 203 provides the engineering specification, defining what capabilities the system must possess and what performance benchmarks it must meet. GB/Z 203 normatively references GB/Z 204, meaning its technical requirements are premised on that architecture. Together, the two documents provide a horizontal framework and a vertical capability baseline.

To conclude, GB/Z 204 and 203 address the immediate pain point of cross-platform fragmentation in industrial foundation models, while GB/Z 195 defines how agents interact across physical and digital environments. These two areas form the core infrastructure and entry points for industrial AI, which is why they were prioritized as flexible guiding standards over more rigid mandates. European stakeholders should treat these as technical references, benchmark them against existing frameworks, and engage through technical dialogue rather than strategic speculation.

Source: https://www.samr.gov.cn/bzjss/tzgg/art/2026/art_ff8536434b844d9b987695e17887d426.html