ENISA and JRC have started testing Chinese open-weight models, three officials confirm

Two EU bodies have started testing Chinese open-weight AI models. The EU Agency for Cybersecurity (ENISA) and the European Commission's Joint Research Centre (JRC), the Commission's in-house science service, are evaluating such systems, Politico reported. Three EU officials confirmed the information. Neither body had previously discussed the work publicly.

The news itself is brief, but it shows how Brussels' attitude toward Chinese AI is changing. It was previously assessed mainly through bans and warnings. Now it is starting to be studied as a technical object.

What is known about the testing

According to the available portion of the material, the picture is as follows:

• two bodies are doing the testing: ENISA and the Joint Research Centre;
• the subject of the testing: open models from Chinese developers;
• sources: a Politico report and three EU officials who confirmed the work;
• status: this is the first public acknowledgment of such work by both bodies;
• not disclosed: the list of models, the methodology, the timeline, and how the results will be used.

The last point matters. It is still unclear whether this is a security assessment, a comparison of capabilities with Western models, or preparation of recommendations for EU institutions. Without these details, any conclusions about political consequences remain guesswork.

Why open-weight in particular

Open-weight models are published in full: you can download the weights, deploy them on your own hardware and take them apart. For a regulator or researcher, this is a fundamental difference from a closed API. A model can be run in an isolated environment and put through your own test suites. You can look for undesirable behavior, check resistance to jailbreaks, see how the model responds to politically sensitive topics, and assess risks when it is used in agentic scenarios.

With someone else's cloud service, this kind of testing is not possible. Data leaves your perimeter, and the model's inner workings are not visible. With open weights, the question of trust in the weights themselves and in the ecosystem around them remains, but it can be investigated empirically.

Chinese labs are actively releasing such models. Developers around the world, including in Europe, use them widely because they can be fine-tuned and hosted in-house. For ENISA, which is responsible for cyber resilience, and the JRC, which advises the European Commission on scientific matters, this is a natural subject of interest.

Context for the European market

European regulators have already dealt with Chinese AI. Italy's data protection regulator restricted DeepSeek in early 2025, and German authorities demanded that Apple and Google remove its app from their stores. Those cases concerned mainly services and the processing of personal data. Testing open weights belongs to a different layer: what matters here is not the fate of an app but the properties of the model itself.

The EU also has the AI Act, which imposes obligations on providers of general-purpose models. How the data collected by ENISA and the JRC might connect to its application has not yet been reported.

What teams already using these models should do

If you build products on Chinese open-weight models, do not wait for Brussels' conclusions. There are several things worth doing now:

• Check each model's license, including restrictions on commercial use and territory.
• Run the model yourself or in a trusted cloud. That way data does not go to the developer, and you control the version.
• Build your own security and quality test suite. It will be useful in any external audit.
• Pin weight versions and download sources. This will simplify verification if ENISA or JRC recommendations appear.
• Follow publications from both bodies. If their findings become public, they will quickly become a reference point for procurement and compliance teams.

For now, the main fact is that the EU has moved from reacting to Chinese AI to studying it. The results of this work will determine how strict, or conversely how practical, future requirements for using such models in Europe turn out to be.