The numbers: filing and registration through 2025
On 9 January 2026 the CAC announced the following:
| Item | New in 2025 | Cumulative total |
|---|---|---|
| Generative AI services that completed filing | 446 | 748 |
| Generative AI applications or functions that completed registration | 330 | 435 |
The implication is plain: a large number of AI products have already completed compliance, and the regulatory risk facing products that launch without filing continues to rise.
Three compliance routes
AI companies must choose the route that matches their technical architecture and business model. The three routes are not interchangeable, and in some cases more than one applies.
Route one: algorithm filing
Applies to: application-layer products that provide algorithmic recommendation services to the public
Legal basis: Article 24, Provisions on the Administration of Internet Information Service Algorithmic Recommendations
A provider of algorithmic recommendation services with public-opinion attributes or the capacity for social mobilisation shall, within ten working days from the date on which it begins providing services, file through the internet information service algorithm filing system the name of the provider, the form of service, the field of application, the type of algorithm, the algorithm self-assessment report, the content proposed for publication, and other information, and complete the filing formalities.
The test: does your product provide algorithm services to the public? Algorithm filing is required wherever any of the following is offered to unspecified users:
- Generative synthesis (AI-generated text, images, audio or video)
- Personalised push (content or product recommendation)
- Ranking and selection (search result ranking, league tables)
- Retrieval and filtering (information retrieval, content filtering)
- Scheduling and decision-making (dispatch, pricing, decision support)
Filing system: https://beian.cac.gov.cn
Review period: thirty working days by statute; two to four weeks in practice where matters proceed smoothly
Route two: large-model filing
Applies to: companies that self-develop, deeply fine-tune or continue pre-training a model foundation, and then provide generative AI services to the public
Legal basis: Article 17, Interim Measures for the Administration of Generative Artificial Intelligence Services
A provider of generative artificial intelligence services with public-opinion attributes or the capacity for social mobilisation shall conduct a security assessment in accordance with relevant state provisions, and shall complete algorithm filing and the formalities for change and cancellation of filing in accordance with the Provisions on the Administration of Internet Information Service Algorithmic Recommendations.
The test: have you trained or modified a model foundation? Large-model filing is required in the following cases:
- Training a large model from scratch
- Continued pre-training on top of an open-source model
- Full-parameter fine-tuning on top of an open-source model
- Modifying the model’s underlying parameters or architecture
How to submit: documents submitted offline to the provincial CAC office with jurisdiction
Review period: practice accounts generally cite six to eight months
Additional documents, compared with algorithm filing:
- Security assessment report
- Model service agreement
- Corpus labelling rules
- List of interception keywords
- Assessment test question set
Route three: large-model registration
Applies to: applications that only call an already-filed third-party large model through its API, without modifying the model’s underlying parameters
The test: do you only call the API of an already-filed model, with no modification at the model layer? Registration applies in the following cases:
- Calling the large-model API of a cloud provider (for example Tongyi Qianwen, ERNIE or Zhipu GLM, each of which is already filed)
- Building application-layer functions on top of that API (chat interfaces, knowledge-base question answering, document processing)
- No training, fine-tuning or parameter modification of the model itself
How to submit: registration with the provincial CAC authority with jurisdiction
Review period: as determined by the local CAC authority
Compliance duties in three typical scenarios
Scenario one: building an AI product on a cloud provider’s large-model API
Company profile: an API from an already-filed large model such as Tongyi Qianwen or ERNIE is used to build a public-facing AI chat or content-generation product.
Compliance duties:
| Item | Required | Notes |
|---|---|---|
| Algorithm filing | Yes | The application layer provides algorithm services to the public |
| Large-model filing | No | The model foundation has not been trained or modified |
| Large-model registration | Yes | An already-filed model API is being called |
Scenario two: serving the public after deep fine-tuning of an open-source model
Company profile: an open-source model such as LLaMA or Qwen is fine-tuned on all parameters using the company’s own data, deployed as its own large model, and offered as an API or chat service.
Compliance duties:
| Item | Required | Notes |
|---|---|---|
| Algorithm filing | Yes | Algorithm services are provided to the public |
| Large-model filing | Yes | The model foundation has been deeply fine-tuned |
| Large-model registration | No | This is not merely a call to a third-party API |
Scenario three: AI capability used only inside the company
Company profile: a large model is deployed internally for knowledge-base question answering, coding assistance or document processing, and is not opened to external users.
Compliance duties:
| Item | Required | Notes |
|---|---|---|
| Algorithm filing | Generally no | No algorithm services are provided to the public |
| Large-model filing | Generally no | No generative AI services are provided to the public |
| Large-model registration | Generally no | As above |
Note: once AI capability is extended from internal use to external customers or ordinary internet users, compliance duties should be reassessed.
Comparison of the three routes
| Dimension | Algorithm filing | Large-model filing | Large-model registration |
|---|---|---|---|
| Core regulation | Provisions on Algorithmic Recommendations | Interim Measures for Generative AI | Supporting guidance under the Interim Measures |
| Regulated object | Algorithm services provided to the public | The model foundation itself | Applications calling an already-filed model |
| Applies to | Applications providing algorithm services | Companies self-developing or fine-tuning models | Applications that only call an API |
| How to submit | Online system at beian.cac.gov.cn | Offline submission to the provincial CAC office | Registration with the provincial CAC office |
| Review level | National and provincial CAC authorities | Provincial first review, national final review | Provincial CAC authority |
| Statutory period | Thirty working days | Six to eight months in practice | As determined locally |
| Core documents | Algorithm description, self-assessment report | Security assessment, corpus rules, interception word list | Authorisation proof, foundation filing number, call chain |
The five principles that decide the route
- A cloud provider’s foundation filing does not make your product compliant — where an already-filed model’s API is called, the application provider must still complete algorithm filing and large-model registration
- Algorithm filing turns on the application layer — does your product provide algorithm services to the public?
- Large-model filing turns on the model layer — have you trained or modified the model foundation?
- Registration turns on the manner of use — do you only call the API of an already-filed model?
- The three are not simple substitutes — some business models require more than one at the same time
Enforcement where filing or registration was absent
Case one: the Chongqing “wrapped AI” penalties (2024) Two AI question-and-answer tools, “Lingxiang Zhiwen AI” and “Shaobing Tuozhan Mi”, operated without any large-model filing or registration and called overseas model APIs directly to serve domestic users for a fee. Each was fined RMB 30,000 and RMB 50,000 respectively, ordered to shut down its official domain permanently, and the operating companies were placed on the CAC’s priority supervision list.
Case two: the Shanghai “Liangjian Pujiang” enforcement action (2025) Three generative AI service websites failed to conduct security assessments, resulting in content such as “face swapping and voice changing”. This was the first case in which the Interim Measures for the Administration of Generative AI Services were applied as the basis for penalties.
Case three: removal of a voice-cloning application (2024) “Reecho”, an AI voice-cloning tool, had no filing or registration record and displayed no labelling for AI-generated content. It was removed from app stores nationwide, with a deadline imposed for rectification.
Frequently asked questions
Q: We call an overseas large-model API such as OpenAI. Is filing required? A: Yes. Where an overseas model API is used to serve the domestic public, algorithm filing and large-model registration are required, and a security assessment for cross-border data transfer may also be engaged. In the Chongqing case, direct calls to an overseas API attracted fines of RMB 30,000 to 50,000.
Q: Our model was only fine-tuned with LoRA. Does that count as “deep fine-tuning”? A: Whether LoRA fine-tuning amounts to deep fine-tuning requires case-by-case assessment. Where only light parameter adaptation is involved (for example a low LoRA rank) and the model’s core behaviour is unchanged, registration rather than filing may apply. Early consultation with the local CAC office is advisable.
Q: Can filing and registration proceed at the same time? A: Yes. Algorithm filing and large-model filing or registration may run in parallel; neither is a precondition for the other.
Q: We have closed our seed round, and our investor says filing must be complete before launch. What should we do? A: Begin filing preparation at the product development stage and run it in parallel with development. The algorithm security self-assessment report usually takes two to four weeks to prepare and is the most time-consuming part of the process. See the complete guide to algorithm filing.
This article is prepared on the basis of the Provisions on the Administration of Internet Information Service Algorithmic Recommendations, the Interim Measures for the Administration of Generative Artificial Intelligence Services, and the CAC’s filing and registration announcements of January 2026. Specific compliance duties should follow the latest position of the CAC authority with jurisdiction. For advice on large-model filing or registration, please contact Geason Law Offices.