China's Cyberspace Administration (CAC) publishes generative-AI filing figures in periodic batches, and the headline numbers tend to travel farther than the rules that produce them. In its July–August 2026 disclosure, CAC and other departments reported filing 124 new generative-AI services nationally, while 133 generative-AI applications and functions completed local registration. Cumulatively, by August 31, 2026, 1,112 services had completed national filing and 731 applications or functions had completed registration.
The more instructive story is not the size of those totals but their shape. China operates a two-tier compliance structure: large models and full services sit in one tier, and the applications built on top of them sit in another. The ratio between the tiers, the direction it is moving, and the newest category of filings — models that run on the phone itself — say more about how the regime is maturing than any single batch of numbers.
Two Tiers, Two Processes
The distinction is procedural, not semantic. Under the Interim Measures for the Management of Generative AI Services, providers of generative-AI services with public-opinion or social-mobilization attributes must file or register through their local cyberspace authorities. CAC's notices then split this into two tracks: national filing of services themselves, and local registration of applications and functions that call already-filed models through an API interface or other means. A model gets filed; an app that leans on that model gets registered.
The design is deliberate. Regulators review the underlying model once, then attach downstream products to it through a lighter local process. CAC's July–August notice restates a disclosure obligation as well: online applications must publicly display which filed or registered service they use, naming the model and its filing or launch number. The registry is meant to be legible to users, not only to officials.
Reading the Ratio
Set the tiers side by side and the cumulative picture is roughly 1.5 filed services for every registered application (1,112 against 731). That gap is what one expects from a regime that formalizes models before the products layered on them: the base layer was addressed first, and the application layer is filling in behind it.
The marginal numbers point the other way. In July–August alone, 124 new services were filed while 133 new applications or functions were registered — so in that window, app registrations outnumbered new model filings. If the pattern holds, the two-tier system is being tested less on the model side, where activity is steady, than on the application side, where growth now runs ahead of the filings meant to anchor it.
Two caveats apply. CAC publishes no denominator for how many generative-AI products exist in China, so the ratio describes what has been filed, not what is on the market. And because the headline totals are cumulative, a single batch cannot establish a trend on its own. The inversion in the marginal ratio is nonetheless a concrete, checkable change to watch in future disclosures.
The On-Device Frontier
Within the 124 services filed in July–August, seven provide on-device generative-AI capabilities on mobile phones — a category that had not previously occupied space in CAC's announcements. On September 23, CAC issued a separate notice naming three generative-AI services that provide on-device model capabilities on mobile phones and have completed filing, one of which it identified as "YOYO Claw."
On-device models matter because they move inference to the handset, where it can run without a round trip to a server. That raises questions the server-based filing model was not built around: what exactly is being filed, how an update to a shipped model is tracked, and what disclosure means when the service ships inside hardware. CAC treats these as generative-AI services on the same filing path, counted alongside cloud-based offerings rather than given their own category. The stated rationale is that on-device filing supports innovation and standardized application; the compliance stack, so far, is the existing one.
Framework 3.0 and the Governance Backdrop
The filings arrived against a broader standards effort. On September 14, at the opening of the 2026 National Cybersecurity Week in Jinan, the National Technical Committee for Standardization of Cybersecurity released AI Safety Governance Framework 3.0, following 1.0 in 2024 and 2.0 in 2025. CAC states the framework was drafted under its guidance by the China Cyberspace Research Institute, the CAC Data and Technology Support Center, and other institutions. It retains the core logic of classifying risks, pairing them with technical responses, and applying comprehensive governance, while updating the risk taxonomy and the mitigation and governance measures.
Framework 3.0 is a standards-body document, not a regulation, so it does not itself expand filing obligations. Its relevance to the two-tier system is directional: it is where risk categories are defined before they surface in filing practice and enforcement.
When the Registry Has Teeth
CAC's enforcement disclosures show the filing-registration distinction is not merely administrative. In cases published September 15, CAC described a Jiangsu technology company running two websites that acted as "API relay stations," calling the interfaces of multiple large-model products to offer dialogue and question-answering services without conducting a required security assessment. CAC said the company was ordered to rectify, warned, and had responsible personnel handled strictly, citing the Interim Measures and the security-assessment rules for services with public-opinion or social-mobilization attributes.
Another case, in Sichuan, involved a WeChat mini-program offering AI text dialogue and image generation that failed to add explicit labels to AI-generated content, failed to add implicit labels in file metadata, and did not complete a security assessment; CAC ordered it taken offline, citing the deep-synthesis provisions, the content-labeling measures, and the assessment rules. The API-relay case is the sharper signal for the two-tier regime: it targets the connective layer between filed models and end-user apps — precisely the tier registration is designed to cover.
The Broader Instrument Set
Filing is one instrument among several. At the same Cybersecurity Week, CAC-guided bodies released Chinese Internet Corpus 4.0 — about 120GB of verified Chinese-language data — alongside Cybersecurity Dataset 1.0 and an AIGC Detection Dataset containing 100,000 text samples and 100,000 image samples. The detection dataset addresses the same problem the Sichuan case illustrates: identifying generated content and tracing its provenance.
What to Watch
The revealing variables ahead are structural. Does the recent inversion — more app registrations than new model filings — persist, suggesting the application tier is where the regime's load is concentrating? Does CAC keep folding on-device models into the existing filing path, or begin to treat phone-resident inference as a category with its own requirements? And as Framework 3.0's updated risk taxonomy filters into practice, will filing and enforcement follow it? The totals will keep growing; the more telling questions are about shape and ratio — and whether a registry built for cloud models can stretch to cover a model that ships in your pocket.
Sources: Cyberspace Administration of China, "Announcement on Filed Generative AI Services (July–August 2026)," https://www.cac.gov.cn/2026-09/14/c_1791136833136332.htm; CAC, "Announcement on Three Filed Generative AI Services Providing On-Device Mobile Capabilities," https://www.cac.gov.cn/2026-09/23/c_1791913977404975.htm; CAC, "AI Safety Governance Framework 3.0 Released," https://www.cac.gov.cn/2026-09/14/c_1791137092283345.htm; CAC, "Enforcement Cases in Cybersecurity, Data Security, and Personal Information Protection," https://www.cac.gov.cn/2026-09/15/c_1790876152357946.htm; CAC, "Chinese Internet Corpus 4.0 and Two Other High-Quality Datasets Released," https://www.cac.gov.cn/2026-09/16/c_1791306277101208.htm.