On 14 September 2026 the Cyberspace Administration of China published the filing information for generative AI services for July and August 2026. In those two months 124 services completed national filing, seven of them on-device services for mobile phones, and a further 133 applications or functions completed registration with local cyberspace administrations. As at 31 August the cumulative totals were 1,112 filed services and 731 registered applications or functions.
Taken alone, these figures are a routine periodic disclosure. Read as ratios, however, they reveal a change in the filing system: it is no longer merely a gate to be passed before launch.
The data
| Indicator | July–August 2026 | Cumulative to 31 August 2026 |
|---|---|---|
| Generative AI services completing national filing | 124 new | 1,112 |
| Of which on-device services for mobile phones | 7 | — |
| Generative AI applications or functions completing local registration | 133 new | 731 |
Two comparisons are worth noting. First, the number of new filings is close to the number of new registrations (124 against 133), which indicates that the two categories of participant in the chain — those developing their own models and those building applications on existing models — are now broadly equal in number. Second, the ratio of total registrations (731) to total filings (1,112) shows that the application layer is expanding quickly, and that most participants at that layer choose registration rather than filing to achieve compliance.
Filing and registration: two parallel channels
The announcement is explicit as to the division between the two channels: a party that develops its own model and provides generative AI services to the public goes through national filing; a party that calls the capability of an already-filed model through an API or otherwise to create an application or function registers with the local cyberspace administration.
The distinction appears technical but in practice determines the compliance route for most AI companies. For a small or medium-sized AI company, completing a filing for a large model of its own is unrealistic in both cost and time; the more common and more realistic course is to use the capability of an already-filed model and then register its own application or function locally. The receiving authority, the materials required and the focus of examination differ between the channels, and a wrong choice will render the compliance ineffective. On how the distinction operates in practice, see large-model filing and registration compared.
Why on-device services are counted separately
Of the 124 filings in July and August, seven were on-device services for mobile phones. Nine days later the Cyberspace Administration issued a separate announcement disclosing the filing information for three on-device services, including YOYO Claw.
The two steps show that on-device services have become a category treated distinctly. This reflects the nature of the service: the model runs locally on the device and data in principle does not leave it, but the character of the service — providing generative capability to the public — is unchanged, and it therefore remains within filing. On the scope of the obligation and the division of responsibility between manufacturers, see filing for on-device generative AI services.
Obligations extending into the operational phase
The most easily overlooked element of the data is the disclosure requirement in the announcement: a generative AI application or function that has launched should disclose, in a prominent position or on the product detail page, the filed or registered service it uses, stating the name of the model and the filing number or launch number.
This requirement changes the character of filing. Filing was originally an act at a point in time — obtain the number and the matter is complete. The disclosure requirement extends the obligation into the operational phase: after launch, the product must continue to display the model it uses and its filing information. An application that calls several models must disclose each of them. Compliance is therefore a continuing state that can be verified, not a qualification obtained once.
Given that “non-compliant AI products and applications” were identified as a priority in the “Qinglang” campaign against AI application disorder, the adequacy of disclosure is likely to be a direct check in inspections. The obligation to label generated synthetic content belongs to the same layer; on the materials required, see the generative AI filing materials checklist.
What to verify
For companies already live or about to launch, the data point to three checks: whether the correct channel has been chosen — filing for a self-developed model, registration for an application built on a filed model; whether disclosure has been implemented, in a prominent position or on the product detail page, stating the model name and the filing or launch number; and whether the information remains consistent with reality — whether filing and registration information is updated when the model version or the scope of capability changes.
None of the three is complex, but each must be satisfied at launch and will be verified on an ongoing basis. The maturity of the filing system is evident precisely in requirements of this kind.
For assistance with determining the correct channel, preparing materials, or implementing disclosure and labelling, see the algorithm filing and large-model registration service page. Related reading: filing for on-device generative AI services and large-model filing and registration compared.
This article is prepared on the basis of the Announcement on the Publication of Filing Information for Generative AI Services (July to August 2026) of 14 September 2026 and the Announcement on the Publication of Filing Information for Three Generative AI Services Provided On-Device on Mobile Phones of 23 September 2026, both issued by the Cyberspace Administration of China. It is for reference only and does not constitute legal advice. Specific filing and registration requirements should be confirmed with the local cyberspace administration.