A procurement officer writes "comply with the applicable national standard for AI agents" into a tender. Two weeks later she cannot answer: which document, with what legal force, and starting when. The documents exist. The ambiguity is what they are.
Key takeaways
- On 27 August 2026 the State Administration for Market Regulation and the Standardization Administration approved 60 national standardization guidance technical documents (国家标准化指导性技术文件) in one batch; 39 titles carry the 人工智能 prefix.
- All 60 are GB/Z guidance documents, not mandatory standards. Under the Standardization Law only 强制性标准 (mandatory GB) can be the basis for refusing a product or service; a GB/Z sets direction, not pass marks.
- The batch spans architecture, compute, agents and datasets: GB/Z 242-2026 on agent requirements, GB/Z 246-2026 on compute platforms, GB/Z 237-2026 on trusted datasets.
- China's large-model (大模型) series mixes legal weights in one numbering family: Parts 1 to 5 are GB/T, Part 7 is a GB/Z published the same day.
- A slower track: on 7 August 2026 the same regulators opened comment on 62 proposed standard projects, led by a foundation-model security testing method, on a 15-month cycle.
- Local governments pay for standard-setting. Suzhou offers up to RMB 1 million (≈ US$149K / HK$1.17M) for leading an international standard, RMB 500,000 (≈ US$74K / HK$584K) for a national one; Guizhou pays RMB 500,000 (≈ US$74K / HK$584K) per national AI standard.
Three prefixes that look identical in a tender
GB is a mandatory national standard. Article 25 of the Standardization Law is blunt: products and services that do not meet mandatory standards may not be produced, sold, imported or provided.
GB/T is a recommended national standard. Voluntary to adopt, but binding once cited in a contract or regulation. Article 21 requires recommended standards never to sit below the mandatory ones.
GB/Z is a national standardization guidance technical document. It gives frameworks and methods rather than pass marks, and is not a conformity-assessment basis. These move fastest, carrying the least legal risk.
That last category is where the August batch sat: all 60 are GB/Z, in force the day they were published, with no transition period — which is how one day moves the reference architecture of an industry.
What the batch covers
The 39 AI-titled documents fall into four layers: data and training (GB/Z 218.1-2026 on embodied-intelligence data quality, GB/Z 220 on generation platforms, GB/Z 237 on trusted datasets); serving infrastructure (GB/Z 243-2026 on on-device inference, GB/Z 246-2026 on compute platforms); architecture and evaluation (GB/Z 221, 222, 226 on model-as-a-service, GB/Z 236 and 253 on retrieval-augmented generation); and agents plus verticals (GB/Z 242-2026 on agent requirements, GB/Z 240 on power-sector agents, GB/Z 265-2026 on transport safety).
Two things stand out. The batch is not purely AI — the remaining 21 cover superconducting magnetic levitation, machine vibration and biomimetic image search among others. And it is heavily vertical: a company deploying in power, steel or ports has a named document for its sector, which is procurement-friendly in a way a general framework is not.
The odd one out: a series that mixes weights
The 大模型 series exposes the confusion most clearly. Same serial number, three legal categories:
- Parts 1 to 3 — general requirements, evaluation indicators, service capability maturity — are GB/T, published in 2025.
- Parts 4 and 5, computer vision and multimodal models, are also GB/T, published on 28 September 2026 and not in force until 1 January 2027.
- Part 7, speech models, is a GB/Z, published on 27 August 2026, already in force.
Part 6 is absent from the register — a gap that usually means under development or withdrawal, though the platform does not say which.
So one "series" has parts in force, parts pending and a missing part — and the one already running is the weakest in law. Part 7's drafters include China Mobile and iFlytek.
The slower track: 62 proposed projects
Published the same month on a different clock, the regulators opened comment on 62 proposed standard projects, including 《网络安全技术 人工智能基础模型安全测试方法》 (testing methods for the security of artificial intelligence foundation models).
Its project record is specific. The plan number is 20264773-T-469, it is administered by TC260, the national cybersecurity standardisation technical committee (全国网络安全标准化技术委员会), and the drafters are the China Information Security Evaluation Center, the China Electronics Standardization Institute and the Pujiang National Laboratory. The cycle is 15 months; the comment window closed on 6 September 2026.
The scope is the interesting part. Built on the assessment methodology of GB/T 18336 (ISO/IEC 15408), it would identify 14 categories of threat — training-data poisoning, model theft, backdoors, adversarial examples and prompt injection among them — across seven security dimensions and five assurance dimensions.
Such a method does not exist as a national standard yet. The record says 制定 — in development — and a 15-month cycle means drafting, comment, examination and approval all still sit ahead. Nobody should write "complies with the national standard for foundation-model testing" into a contract.
Honest limitations
I am working from the official register of published standards and draft project records, not the technical texts, which sit behind the platform's full-text access. The 39-of-60 count is my own tally, not an official statistic; the batch size of 60 is official. I have not checked whether any of the 21 non-AI documents carry AI content under a title omitting the 人工智能 prefix, so the AI share is a floor rather than a ceiling. Project cycles are planned durations and slip. And I cannot assess how often a GB/Z document lands in a Chinese procurement document, the only route by which guidance documents gain indirect force.
What readers can do now
- Check the prefix before citing a number. Writing "符合 GB/Z 242-2026" promises something the document cannot adjudicate; "参照 GB/Z 242-2026" is defensible.
- Check the status field, not the announcement date. A GB/Z published in August 2026 is in force; a GB/T published the same month may not be until 2027.
- Read the sector document, not the general one. If you deploy in power, steel or ports, a vertical GB/Z names your use case. If you build agents or run on-device inference, GB/Z 242 and 243 come first.
- Watch the 62-project list too. A completed foundation-model testing method becomes the yardstick for every security claim in a filing dossier.
Why the volume is the story
China's AI standards effort is no longer a blueprint with a target. It is a production line, running in batches, mixing three document types with three legal weights and three effective-date conventions.
The difficulty is not the absence of standards. It is their abundance, and the fact that the fastest-moving layer is the one that cannot settle a dispute.
According to Li An, chief scientist at BrainNet (脑机网), the discipline for reading any Chinese AI standard is to ask two questions first — what kind of document is this, and on what date did it become usable — because a document that exists and one you can be held to are different things.
The count went up by sixty on a Tuesday. The enforceable count went up by one.