In April 2026 the Ministry of Industry and Information Technology (MIIT) and the National Data Administration (NDA) jointly issued a notice on implementing the 2026 “Model-Data Resonance” action, setting seven priority tasks to promote coordinated and mutually reinforcing development of AI models and data resources — a virtuous cycle of “models drawing data, data feeding models”.

What it is meant to solve

The relationship between models and data has long been mismatched: models iterate quickly, but usable high-quality data grows slowly. The result is that model capability is constrained by data supply, while large volumes of dormant data cannot be used effectively.

The idea of “model-data resonance” is two-way. “Models drawing data” means using model needs to pull the collection, labelling and governance of data, giving data production a clear target; “data feeding models” means using high-quality data to improve training and optimisation, raising model performance in specific scenarios. The two authorities acting jointly connects the policy tools on the model side and the data side.

Where the seven tasks land

From public information, the seven tasks centre on data supply, labelling capability, scenario deployment and ecosystem building, covering the chain from data collection to model application. Their common aim is to move data from “available” to “usable”, and models from “general” to “applicable”.

For companies, this means the path to participation is clearer — as data supplier, labelling provider, model party or scenario party, each can find a role — and the entry threshold is clearer too, especially on lawful sourcing and data quality.

Data compliance is the unavoidable prerequisite

Whatever the role, data compliance is a precondition.

First, lawful sourcing. Data used for model training and dataset building must have a lawful source and authorisation basis. Publicly collected, third-party purchased and user-generated data have different compliance foundations and must be reviewed separately — see our data-compliance service.

Second, personal-information boundaries. Where personal information is involved, the purpose, scope and notice-and-consent arrangements must be clear; sensitive personal information is stricter. Those processing personal information at scale should watch the special duties of “large personal information processors” — see our analysis of duties of large personal information processors.

Third, ownership and circulation. The ownership arrangements for datasets and their external circulation should be designed alongside the data property-rights registration system — see our analysis of the draft guide on data property-rights registration.

Fourth, security measures. Storage, transmission and sharing of datasets must be protected in line with data sensitivity.

“Model-data resonance” and the NDA’s earlier deployment on high-quality industry datasets are two advances along the same line. The former focuses on the coordination mechanism between models and data; the latter on the scaled building of datasets — see our analysis of the action plan for high-quality industrial datasets.

For AI companies, the combined signal is clear: data supply is becoming work with explicit policy support and explicit compliance requirements. Getting data source, ownership and security right is both the ticket to participate and the basis for surviving due diligence in future financing and M&A — see our data-compliance service.

This article is based on public information concerning the notice jointly issued by MIIT and the National Data Administration in April 2026 on implementing the 2026 “Model-Data Resonance” action. It is for general reference only and does not constitute legal advice.