The South China Morning Post reported this week that Chinese AI models are quietly entering European markets despite regulatory concern. It is a trend we recognise from our own daily work at ai-jarvis.eu, where we run open-weight models on local hardware as part of our testing. The technical appeal is simple: open weights, low cost and full control over data. The political picture is far more complicated.
Why European developers say yes
The models in question are not consumer chatbots. They are open-weight models — neural networks whose internal parameters are published, so anyone can download them and run them on their own hardware. The two names Europeans encounter most are DeepSeek, from a Chinese research lab, and Alibaba's Qwen family.
DeepSeek's R1 model, released in January 2025, matched leading Western reasoning models while costing a fraction of the amount to train and run — a fact documented widely at the time. That combination of performance and price is hard to ignore, especially for startups and small businesses.
The Qwen family from Alibaba is Apache-licensed and comes in many sizes, from small models that fit on a phone to large ones for server deployment. On open-model platforms, Qwen and DeepSeek variants are among the most downloaded in Europe, and they handle European languages — German, French, Czech — well enough for everyday work.
Why Brussels is uneasy
The concern was never really about the technology itself. It is about what happens when a model — or the data you feed into it — crosses borders.
When you call a Chinese cloud API, your input is processed on servers outside the EU, potentially in mainland China. European regulators have long pointed out that Chinese law, including the 2017 national intelligence law, obliges organisations to cooperate with state intelligence services. That is a genuine question for any European business, not a hypothetical one.
Concrete steps have already been taken. Italy's data protection authority blocked the DeepSeek chatbot in early 2025 over data concerns, and several EU institutions restricted its use on official devices.
What the EU AI Act does — and doesn't — do
The EU AI Act, in force since August 2024 with obligations phased in over the following years, sets rules for general-purpose AI models. The official EU text requires providers to publish technical documentation and a summary of training data, and to respect copyright law.
But there is a carve-out that matters enormously in this story: models released under a free and open-source licence are exempt from most of those obligations unless they qualify as posing systemic risk. The threshold is training compute above 10^25 FLOPs — a level that, for now, covers only the very largest frontier models.
In practice, this means an open-weight Chinese model downloaded and run locally in Prague, Lyon or Warsaw faces almost no direct restriction under the AI Act. The rules bite when a model becomes a large commercial service — not when it runs on your own computer. The Commission built a review mechanism into the AI Act precisely because this area is moving quickly.
Local versus cloud: where your data actually goes
This brings us to the choice European users actually face. It is not a simple choice between "Chinese AI" and "European AI". It is a choice about where your data goes.
Run the model locally — with a tool like Ollama on a laptop or a desktop workstation — and nothing leaves your device. This is a big reason open-weight Chinese models appeal to European developers: local deployment sidesteps GDPR data-transfer issues entirely. For a teacher, a nurse or a small business owner, the same logic holds. A model that runs on your own machine is a privacy-friendly model.
Use the Chinese provider's cloud API instead, and your data travels to servers outside the EU. That is not automatically illegal under GDPR, but the provider must offer appropriate safeguards, and any employer should think hard about what staff are allowed to upload. Many European companies have therefore chosen local deployment of open models — a path where Chinese technology ends up helping Europeans keep data in Europe.
Europe's own answer
European AI is not standing still. France's Mistral builds competitive models and has become the clearest homegrown alternative, with enterprise offerings and cloud partnerships across the EU. The EU is also investing heavily in its own AI infrastructure.
But the honest conclusion is that open weights are a global market. A developer in Tallinn can download a model trained in Hangzhou today, and no law currently stops her. Brussels can shape where data flows and impose duties on the largest providers. It cannot put the open-weight genie back into the bottle.
The real question for Europeans, then, is less about where a model was born and more about where their data sleeps. On that front, the new open-weight models from China have become unexpected allies: because they run locally, they offer a level of control that commercial cloud services rarely do.
Is it legal to use Chinese AI models in the EU?
Yes, in most cases. Open-weight models like DeepSeek and Qwen that you run locally fall under the open-source exemption of the EU AI Act unless a model is classified as posing systemic risk. Using a Chinese cloud API is also legal, but GDPR obligations apply to the data you send.
Can Brussels ban these models?
No general ban exists and none is planned under the AI Act, which works through obligations on providers and high-risk uses rather than outright bans. Individual EU countries can restrict specific services on data-protection grounds, as Italy did with the DeepSeek chatbot.
How do I know a model is safe for my data?
Check where the data goes. A model running locally with a tool like Ollama keeps everything on your device. With a cloud API, read the data-processing agreement, check whether servers are located in the EU, and avoid uploading personal or confidential information.
