A state-owned utility wants to buy a large model (大模型) to run its maintenance desks. Its procurement team opens three vendor pitches and finds three different ways of saying "we're the best," with no common test any of them can repeat. The gap between marketing and measurement is exactly what a new national standard tries to close.
China now has that standard. It is not a law and not a ban. It is a three-part technical specification that tells model builders and buyers what "capable, safe, and service-ready" should mean in concrete terms.
The document, and what stage it is at
The GB/T 45288 series, "Artificial Intelligence — Large Model (人工智能 大模型)", is a recommended national standard (推荐性国家标准) — the "GB/T" prefix signals voluntary adoption, not a mandatory compliance line. This distinction matters and is easy to miss: unlike the mandatory GB 45438-2025 that governs AI-content labeling, GB/T 45288 is a yardstick companies can choose to measure against.
It is not a draft or a proposal. According to the national standards information platform run by the Standardization Administration, all three parts have been published and are in force (现行):
- Part 1, General Requirements (通用要求) — GB/T 45288.1-2025, published and effective 28 February 2025.
- Part 2, Evaluation Indicators and Methods (评测指标与方法) — GB/T 45288.2-2025, published and effective 28 February 2025.
- Part 3, Service Capability Maturity Assessment (服务能力成熟度评估) — GB/T 45288.3-2025, published and effective 24 January 2025.
The series is described by state media as the country's first national standard focused on general large models, filling a blank in how model capability is evaluated.
Three parts, three jobs
The split is deliberate.
- Part 1 sets what a model must cover: performance, safety, and the basics of how it should behave. It is the "what good looks like" definition.
- Part 2 sets how you measure it — the indicators and the test methods, so two labs running the same model should reach comparable conclusions.
- Part 3 steps back to the provider: it grades the maturity of the organization's service capability, not just the model on a benchmark.
Together they move the conversation from "trust our demo" to "here is the test we ran, and here is the grade."
What it actually measures
The standard spans three axes that buyers care about:
- Capability — how well the model handles the tasks it is pitched for, including accuracy and the tendency to hallucinate.
- Safety — content safety and the controls a provider puts around generation, including how risks are monitored as the model updates.
- Service readiness — stability, update discipline, and the operational maturity of the team behind the model.
None of this is abstract. The framing is built for procurers: a company or agency that wants a model can point at Part 2 and ask a vendor to submit to the same evaluation indicators.
Who it is for
The stated audience is domestic model developers, downstream integrators, and public-sector buyers. In practice the most visible pull has come from state-owned enterprises. Reporting by state media describes the standard's supporting evaluation system being used to build an authoritative shortlist that guides central and state-owned enterprises in model selection, and to support industry models at firms such as Sinopec (中石化) and China Southern Power Grid (南方电网), with scenario validation carried out on models including Huawei's Pangu (华为盘古), iFlytek's Spark (讯飞星火), and China Mobile's Jiutian (中移动九天).
A companion benchmark, the "Qiuzuo"-LMBench evaluation baseline, has been described as aligned with the standard and recognized by the China National Accreditation Service for Conformity Assessment (CNAS). These specifics come from state-media reporting rather than the standard text itself, so they should be read as the ecosystem growing around the standard, not as clauses of it.
Why a voluntary standard still matters
A recommended standard has soft power. Once a large buyer — especially a state-owned one — adopts it as a procurement filter, vendors follow whether or not the law compels them. The same pattern appears in product safety and green ratings worldwide: the spec becomes the door.
There is also an outward dimension. State media report that the standard's materials are being prepared in English, Lao, and Khmer for the ASEAN region, an early sign that China's model-evaluation language may travel with its open models.
Honest limitations
- It is voluntary. GB/T means recommended, not mandatory. A vendor can ignore it without breaking a law — though it may struggle to sell to compliance-driven buyers.
- It measures, it does not license. Passing the standard is not a safety clearance under the Interim Measures for Generative AI Services (生成式人工智能服务管理暂行办法) or the labeling rules. The two regimes sit beside each other.
- Benchmarks age fast. A standard finalized in early 2025 will lag models released a year later; the evaluation indicators need revision to stay meaningful.
- Quantitative claims about adoption are reported, not statutory. Figures on evaluation tasks run or vendors assisted come from media, not the standard's text, and should be treated as indicative.
What readers can do now
- If you buy or integrate models in China, cite GB/T 45288 Part 2 in procurement so vendors face one test, not three self-authored ones — and ask for the CNAS-aligned benchmark result, not a marketing sheet.
- If you build models, run your own release against Part 1 and Part 3 before talking to state-owned or enterprise buyers; the maturity grade often matters as much as the score.
- If you track governance, watch whether this recommended standard hardens into a procurement requirement — that is where its real bite will come from.