The three routes are not three parallel options but the result of running a single sequence of questions. The most common error is to ask first how long it takes and how much it costs, and only then which route applies. Once the order is reversed, the documents usually have to be prepared again from the beginning.
The distinction in one line: they govern different things
| Point of comparison | Algorithm filing | Large-model filing |
|---|---|---|
| What it governs | How content is selected and pushed — recommendation, ranking, retrieval and scheduling mechanisms | Whether the model and corpus are safe — whether the model is self-developed or substantially modified |
| Who must comply | Providers of algorithmic recommendation services with public-opinion attributes or the capacity for social mobilisation | Providers of generative AI services to the domestic public that self-develop or substantially modify a model |
| Core documents | Algorithm information, algorithm security self-assessment report, content proposed for publication | Model and corpus information, security assessment report, service agreement, complaints and minor-protection measures |
| Submission | Online filing through the algorithm filing system | Offline submission to the provincial CAC office, with provincial review followed by national review |
| Period | No statutory deadline; months in practice | Usually substantially longer than algorithm filing; quarterly planning is safer |
| Substitution | Not permitted | Not permitted |
The conclusion is a single sentence: the two rest on different instruments and have different triggers; they are not alternatives. Where a product combines model generation with a feed or ranking mechanism, both duties arise, and completing one leaves the other outstanding.
The outcome first: where five situations fall
| Your situation | Route | Also required |
|---|---|---|
| Self-trained or substantially modified model, offered to the domestic public | Large-model filing | Security assessment; generative-content labelling; algorithm filing where feed distribution is present |
| Calls an already-filed model’s interface without altering its capability | Large-model registration | Labelling duty; algorithm filing where feed distribution is present |
| Not offered to the public, but involves recommendation, ranking, retrieval or scheduling decisions | Algorithm filing | Deep synthesis filing assessed as applicable |
| Internal use only, with no recommendation distribution and no external generation | Filing generally not triggered | Retain records of data provenance and content compliance self-inspection |
| Face generation, voice cloning or virtual-avatar driving | Deep synthesis service filing | Labelling duty and content review mechanism |
The first question: have you modified the model
The starting point is not the business model but the provenance of the model. Self-training, continued pre-training, large-scale instruction fine-tuning on a company’s own data and replacement of the foundation model all change the model’s capability boundary and fall on the filing side. Prompt engineering, retrieval augmentation, interface wrapping, or light fine-tuning of tone and format that does not change the range of capability usually fall on the registration side.
This line is not the only criterion written into law, and provincial CAC offices differ in how they assess “substantial modification”. It is not unusual for the same technical solution to produce different conclusions in two provinces. The prudent course is to settle classification first: before preparing documents, discuss model provenance, training method and corpus composition with the local CAC office, confirm which route applies, and only then address the documents.
The distinction between self-development and mere calling is set out in more detail in Large-Model Filing versus Registration: Three Compliance Routes and How They Apply, which should be read alongside this article.
Where the three routes diverge
| Dimension | Algorithm filing | Large-model filing | Large-model registration |
|---|---|---|---|
| Applies to | Algorithmic recommendation services with public-opinion attributes or social mobilisation capacity | Self-developed or substantially modified models offered to the public as generative AI services | Applications calling an already-filed model without altering the model itself |
| Legal basis | Article 24, Provisions on the Administration of Internet Information Service Algorithmic Recommendations | Article 17, Interim Measures for the Administration of Generative Artificial Intelligence Services | The preceding article, together with the rolling mechanism of filing announcements |
| Submission | Online through the algorithm filing system | Offline to the provincial CAC office, with provincial review followed by national review | Registration with the provincial CAC office |
| Core documents | Entity information, algorithm information, self-assessment report, content proposed for publication | Model and corpus information, security assessment report, service agreement, complaints and minor-protection measures | Entity and product information, the filing number and authorisation of the model called, data processing and content security measures |
| Review | Primarily formal review, followed by publication on the filing list | Document review and security assessment together, at a higher level | Primarily verification of the calling relationship, with lighter documentation |
| Period | No statutory deadline; months in practice | No statutory deadline; usually substantially longer than algorithm filing, so quarterly planning is safer | The shortest, though still dependent on local processing |
| Common returns | Self-assessment report inconsistent with the algorithm’s actual function; missing content proposed for publication | Inadequate corpus provenance and authorisation; templated security assessment; fictitious complaints channels | Filed under the wrong category; calling relationship and filing number that cannot be matched |
Read this table as a single line: settle model provenance, then service recipients, then whether public-opinion attributes are present, and finally whether there is an independent distribution mechanism. Once the four questions are answered the route converges, and only the documents remain.
Why large-model filing does not make algorithm filing unnecessary
The two rest on different instruments and have different triggers; there is no substitution between them.
Algorithm filing governs the information distribution mechanism. Article 24 of the Provisions on the Administration of Internet Information Service Algorithmic Recommendations requires providers of algorithmic recommendation services with public-opinion attributes or the capacity for social mobilisation to file, within ten working days from the date services begin, the provider’s name, the form of service, the field of application, the type of algorithm, the self-assessment report and the content proposed for publication. What it addresses is how content is selected and to whom it is pushed.
Large-model filing governs the safety of the model and its corpus. Article 17 of the Interim Measures for the Administration of Generative Artificial Intelligence Services requires generative AI services with public-opinion attributes or the capacity for social mobilisation to conduct a security assessment under relevant state provisions and to complete algorithm filing formalities. In practice this duty takes the form of submitting model, corpus and security assessment materials to the provincial CAC office, which are then reviewed and entered on the filed list.
Where a product combines model generation with a feed, a ranking table or retrieval ranking, both duties arise at once. Completing one leaves the other outstanding.
Deep synthesis is a third regime. The Provisions on the Administration of Deep Synthesis of Internet Information Services impose separate filing requirements for face generation, voice cloning, virtual-avatar driving and similar functions. The Measures for the Labelling of AI-Generated Synthetic Content, in force from 1 September 2025, then isolates the labelling duty: both explicit labels and implicit labels in file metadata must be implemented, and the absence of a label is itself a breach, irrespective of whether filing has been completed. This layer is often overlooked, yet it is the one most likely to surface in spot checks.
Five real scenarios and where they fall
A vertical application calling an already-filed model. Products for contract review, customer service question answering or document summarisation draw their capability from a filed interface and do no training themselves. They fall on registration. Where the product also has a “you may also like” feed or an algorithmically ranked recommendation slot, algorithm filing arises as well.
An open-source model lightly fine-tuned before launch. Whether instruction fine-tuning of an open-source foundation on a company’s own data amounts to substantial modification depends on whether it changes the model’s knowledge boundary and capability range. In practice most provinces treat this as filing. Corpus provenance and authorisation should be prepared in advance, as this is the most common amendment point in this category.
A self-developed model opened to the public. Filing, security assessment and content labelling proceed together. Where user posting, comments or a gallery of works are also opened, algorithm filing arises separately; where minors are involved, minor-protection measures must be implemented in the product interface rather than left in policy documents.
An internal corporate knowledge base. Used only by employees and not offered externally, it generally does not trigger filing. But once that question-answering function is connected to the corporate website’s customer service or an external mini-program, the service recipients become the public and the assessment changes.
AI short drama and digital human broadcasting. Deep synthesis filing is the main line, with the labelling duty following; where the generation stage uses a self-developed or substantially modified model, filing is added on top. Three duties run in parallel, and the absence of any one may cause the product to be blocked at platform level.
Returns that could have been avoided
A corpus that cannot be explained. Stating only that data “comes from the public internet” is insufficient. Collection method, authorisation chain, whether overseas corpora are included and how personal information is handled must all map to specific documents.
A templated security assessment. Risk items in the report that do not correspond to the product’s actual functions are the most common ground for return. The assessment must state what this product generates, where errors may arise, and what measures address them.
A misclassified category. An application calling an interface that files as a model provider, or a self-developed model filed as a registration, will be returned. Classification should be settled before submission, not left until an amendment notice arrives.
A fictitious complaints channel. A telephone number that no one answers, or an email address that no one monitors, is a substantive failure rather than a formal defect.
Labelling that was never implemented. Generated content carries no explicit label, and exported files carry no implicit label in their metadata. Since the Labelling Measures took effect, this item has moved from recommendation to mandatory check.
Official sources
- Provisions on the Administration of Internet Information Service Algorithmic Recommendations, in force 1 March 2022: https://www.gov.cn/zhengce/zhengceku/2022-01/04/content_5666429.htm
- Interim Measures for the Administration of Generative Artificial Intelligence Services, in force 15 August 2023: https://www.gov.cn/zhengce/zhengceku/202307/content_6891752.htm
- Provisions on the Administration of Deep Synthesis of Internet Information Services, in force 10 January 2023: https://www.gov.cn/zhengce/zhengceku/2022-12/12/content_5731431.htm
- Measures for the Labelling of AI-Generated Synthetic Content, in force 1 September 2025: https://www.gov.cn/zhengce/zhengceku/202503/content_7014286.htm
- Internet information service algorithm filing system: https://beian.cac.gov.cn
Further reading on this site: A Complete Guide to Algorithm Filing, Checklist of Documents for Generative AI Filing, Guide to Deep Synthesis Service Filing, Key Points of the Measures on Labelling Generative Synthetic Content, and the overall framework of AI legal services.
Last updated: 27 October 2026. Where regulations or local practice change, the latest official publication prevails.