Open language models now run in the Netherlands too

Together with Nebul, we run large open language models entirely in the Netherlands. For the first time, every part of vragen.ai can run on Dutch soil: the application, the knowledge base and the language model. Why that matters now, and ten questions with answers.

Bing van Moorsel Co-founder, CEO

28 August 2026 10 min read

Ask an AI assistant a question and something invisible happens. Your question travels to a data center, a language model computes an answer and that answer travels back. Where that data center stands and who controls that model is something most organizations only consider once someone asks.

That question now comes from every direction. Boards, data protection officers and procurement teams want to know who their digital work depends on. Not without reason: Dutch public broadcaster NOS examined 16,500 domain names of Dutch governments, healthcare institutions, schools and vital companies early this year, and 67 percent turned out to depend on at least one American cloud service, nearly half on Microsoft alone. Rarely did anyone actually choose that.

For vragen.ai, we now have a new answer to that question. Together with Nebul, we run large open language models entirely in the Netherlands. For the first time, every part of vragen.ai can run on Dutch soil: the application, the knowledge base and the language model that writes the answer.

That delivers on something we have been working toward for a while.

"I really want to play a part in offering sovereign AI, AI from Dutch soil, where you decide from start to finish what comes out of a chatbot."

Matthias, founder of vragen.ai and owner of SWIS

For a long time, that was only possible by giving up quality. That has changed, and that is why this is the moment.

Two thirds of the Netherlands leans on American cloud

Digital dependency was a policy-paper topic for years. In May 2025 it got a face: the chief prosecutor of the International Criminal Court in The Hague lost access to his Microsoft mailbox after American sanctions and moved to a Swiss service. Accounts differ on who exactly pulled the plug, but the lesson stuck: whoever works on an American service can also lose that service.

The numbers behind it are not small. Amazon, Microsoft and Google together hold around 70 percent of the European cloud market; all European providers combined sit at 15 percent, down from 29 percent in 2017. The Dutch parliament now wants the government to speed up work on Dutch alternatives.

To the visitor of your website, all of this is invisible. To the people responsible for data in your organization, it is not. They ask exactly the question this article is about: where does it run, and who controls it?

Where a model runs and who controls it

vragen.ai was built to be model-independent from the start. For every agent you choose which model runs the conversation, and every model carries a label: NL, EU or US. Since this summer, that gives you three flavors.

Where the model runs	Examples	Who controls the model	What answers your question
Netherlands, at Nebul	open models such as Kimi, Qwen, GLM and Mistral	we do: we run our own copy	our own infrastructure in the Netherlands
EU, at Microsoft	most GPT models	OpenAI	a Microsoft service in a European data center
United States	Claude and some GPT variants	Anthropic and OpenAI	the provider's service in the United States

We use the GPT models through Microsoft in European data centers. Processing stays inside Europe, with data processing agreements in place and no training on your data. But the model remains a service: the maker decides which versions exist and when a version disappears.

The models at Nebul work differently. We hold the copy we run ourselves. Answering a question needs no connection to the maker: nothing leaves, and the maker sees no conversation. If the maker ever stops developing a model, our copy does not disappear. Moving on becomes a choice we plan and test ourselves.

GPT and Claude remain available. The Dutch models are added to the list, so you can weigh what fits per use case.

The best open models compete at the top

For a long time the trade-off was simple: if you wanted the best model, you ended up with the large American providers. That is over, and that is not a feeling but measurement (as of August 2026).

Independent benchmarking firm Artificial Analysis ranks the latest version of Kimi, an open model from China's Moonshot AI, first among more than a hundred measured open models. On release this summer, that same model ranked third worldwide, open and closed combined, comparable to the OpenAI and Anthropic flagships of that moment. Research institute Epoch AI calculated this spring that the best open models trail the closed top by four months on average.

Four months. That is the entire distance. And in our own testing, answering questions from a knowledge base with source references, that distance is rarely what makes an answer better or worse; the quality of the source weighs heavier.

Which models are in vragen.ai right now, and how they score on smartness and speed, is shown in the model picker in the product. New versions follow each other monthly, so that list changes faster than an article can keep up with. That is why we deliberately name no version numbers here: the current state lives in the product, this article explains how to choose.

"But those are Chinese models"

Kimi comes from Moonshot AI in China, Qwen from Alibaba and GLM from China's Zhipu AI. That understandably raises questions.

Which is why the distinction between origin and processing matters so much. The model file sits at Nebul in the Netherlands. Answering a question needs no connection to Moonshot, Alibaba or Zhipu. No question travels to China and the maker receives no conversation content. If a resident of a municipality asks a question that Kimi answers, that processing stays in the Netherlands.

The origin does not go away. The model was developed and trained in China, and we had no part in that. But for day-to-day use, we do not depend on the maker. We find that distinction more relevant than the country behind the name.

Nebul can do what only a handful of parties in Europe can

Nebul is a Dutch AI cloud company from Leiden, the same city as SWIS, the agency behind vragen.ai. Since 2022 it has been building compute clusters on NVIDIA hardware in Dutch data centers (source in Dutch). Research firm Gartner placed Nebul on its worldwide list of Cool Vendors among specialty AI cloud providers in 2025, and software company Centric picked Nebul this spring as the foundation for AI services to Dutch government (source in Dutch).

That so few parties can do this has a sober reason: size. The model file of a large open model is well over a terabyte. Getting fast answers out of it means spreading that file across dozens of specialized chips working as one machine; a documented production setup for the latest Kimi uses sixteen of them with over two terabytes of memory combined. Machines like that are scarce, and in the Netherlands almost nowhere available as a service.

In Europe, the number of providers serving models of this size can be counted on one hand. In the Netherlands, we found exactly one that does it on Dutch soil. And look closely at the difference between a Dutch headquarters and Dutch servers: there are providers with an Amsterdam address whose GPUs stand in Finland.

Two parties in the Dutch chain

For the models running in the Netherlands, the operational chain is short. The application and the knowledge base sit with TrueFullstaq in Amsterdam, our hosting partner since 2013. The language model runs at Nebul. You are the controller for your visitors' data; we are your processor.

A question goes to the model, the model computes the answer and sends it back. Nebul does not store the content, does not log it and does not use it for training.

Within vragen.ai we do keep the conversation, so you as an editor can read back what visitors ask and where your content falls short. If someone types personal data into a question, we filter it out within an hour. Conversations are kept for twelve months by default; you can shorten that period.

Our page on privacy, certifications and hosting describes in more detail which data we process and which parties are involved.

Sovereignty is not an absolute

A language model running in the Netherlands does not make the whole technology Dutch. The hardware contains technology from the United States and Taiwan, among others. Kimi, Qwen and GLM were developed and trained in China, Mistral in France. We had no part in their training data or training process.

On quality, too, no single model is best in every situation. For some tasks the heaviest closed models perform better. In other cases the models in the Netherlands offer comparable quality, with the chain set up differently.

So we are not replacing anything. We add options and show where every model runs, so you can decide which combination fits your own knowledge base and your own visitors.

Ten questions we get about this

These are the questions that come up most often in conversations about this topic. Short answers; the details are above and on the security page.

1. Where does a visitor's question get processed?

Always at vragen.ai in Amsterdam, where the application and your knowledge base sit. The language model is a choice per agent: in the Netherlands, in the EU or in the United States. The label on every model says where it runs.

2. Can the US government reach data held by an American company, even on European servers?

US law can require American companies to hand over data, even when it sits on servers in Europe (source in Dutch). Nobody knows exactly how often that happens; that it can happen is written into the law. If you pick a model that runs in the Netherlands, that part of the chain contains no American company.

3. GPT runs through Microsoft in Europe. Is that not enough?

For many use cases it is: processing stays inside Europe and your data is not used for training. But the model remains a service from an American company, under American law, and the maker decides which versions exist. That is why vragen.ai makes it a choice per agent rather than all or nothing.

4. Are open models as good as GPT or Claude?

The best open models sit at the level of the closed top of roughly four months ago. For answering questions from a knowledge base with source references, they fully hold their own in our testing. The current scores per model are in the product.

5. Some open models come from China. Does our data go to China?

No. The model runs at Nebul in the Netherlands and needs no connection to its maker to answer. No question travels to China and the maker sees no conversations.

6. What does "open" mean for a language model?

The maker publishes the trained model file so others can run it themselves. The training data and training process usually do not come with it, which is why this is also called "open weights". What counts is the consequence: we run our own copy and never need to call the maker.

7. What happens if the maker of a model stops?

With a closed model, the provider decides when a version disappears, and you test again. The version running at Nebul does not disappear. Moving to a new version becomes a planned choice, with time to test.

8. Who can read along with conversations?

With a model in the Netherlands, the chain consists of vragen.ai and Nebul, and Nebul does not store or log the content. You read conversations back in your own environment as an administrator; personal data is filtered out within an hour.

9. Does anyone train on our visitors' questions?

No. Not us, not Nebul, and not the makers of the closed models either; that is contractually agreed. With the models in the Netherlands it is not even possible: the maker never sees the conversations.

10. Does this make vragen.ai one hundred percent sovereign?

No, and be careful with any party that promises it does. The hardware contains American and Taiwanese technology and the models were trained elsewhere. What now sits in the Netherlands: where the model runs, which version we use and who is needed to answer a question.