On 15 July 2026 the Cyberspace Administration of China (CAC) published a notice adding seven on-device generative AI services for mobile handsets to its filing list, including Apple Intelligence. This is another batch of centralised publication following the first on-device filings in the first half of 2026.
The shift of on-device filing from “occasional” to “normalised” is itself worth recording: it means that generative AI capability on handsets is no longer an implicit feature of device makers but a service form brought under filing management and subject to public disclosure.
How on-device filing differs from cloud filing
In a cloud generative AI service the model runs on the provider’s servers, with a clear data flow and a single responsible entity. On-device is different: the model capability is deployed on the terminal, and computation happens in the user’s hands.
Two key differences follow.
First, the division of responsible entities. An on-device service is often delivered by a “model provider plus device maker” together; who is responsible for filing, who for content safety, and who for user notice must be agreed from the outset of the cooperation.
Second, the boundaries of data processing. A selling point of on-device is that data does not leave the device — but that only means it “can” stay, not that it “will”. If an on-device feature needs to call a cloud model or return data for optimisation, the processing logic falls back into the cloud-filing framework.
Three compliance points for on-device products
Data localisation must be delivered technically. Products that claim “data does not leave the device” must reflect this consistently in the privacy policy, system settings and technology. Claiming “local processing” in front-end copy while returning data in the background is a high-risk point for regulators and user complaints. See our Data Compliance service page.
Generative-content labelling must cover the on-device chain. Content generated on-device is also generated and synthesised content, and both explicit and implicit labelling chains must be in place. The peculiarity of on-device is that content may be generated without a network, so the implementation of implicit labelling must be designed in advance. See our Algorithm Filing service page.
Minors protection needs an on-device solution. The handset is the main entry point through which minors encounter AI. Anthropomorphic interaction and emotional-companionship on-device features must implement the requirements on minors mode and reminder mechanisms under the Interim Measures on Anthropomorphic Interactive AI Services.
Three points on the division of duties between makers
First, write the filing duty into the cooperation contract. Between the model provider and the device maker, it should be clear who is the filing entity, who supplies materials, and how updates are synchronised on change.
Second, distinguish “on-device features” from “connected features”. Within a single product, which capabilities are on-device and which are in the cloud must be documented clearly, so that the filing description does not conflict with reality.
Third, watch local registration. Applications that call on already-filed model capabilities through APIs or otherwise need registration with the local cyberspace authority, and hybrid on-device/cloud products are especially prone to this. See our Algorithm Filing service page.
For support on the filing path for on-device products, the division of maker responsibility and contractual arrangements, see our Algorithm Filing and Commercial Contracts services.
This article is based on the on-device generative AI service filing notice published by the Cyberspace Administration of China on 15 July 2026. It is for reference only and does not constitute legal advice.