October 19, 2026 · Triox Informatika · 6 min read
Open and European models—when the data cannot leave the building
Not every AI runs somewhere in America. The world of Mistral, Llama, DeepSeek and downloadable models works by different rules. We look at what an “open” model means, when it is worth running AI on your own computer or server, and what that gives a Hungarian company.
In brief
- An open model can be downloaded and run on your own computer or server. The data does not leave the company, there is no monthly fee per user, and you do not depend on a single provider.
- The best-known open models are Meta’s Llama, the French Mistral, the Chinese DeepSeek and Qwen families, and the smaller open variants from OpenAI and Google.
- Mistral is the largest European player, with its own assistant and EU-based data processing—for those who want the cloud, but in European hands.
- The price of running locally is performance and convenience: open models are good, but they lag behind the top models, and operating them takes expertise. For specific, well-defined tasks, however, they are excellent.
There is a question that comes up at almost every corporate AI meeting: “And where does the data go?” The answer for ChatGPT, Claude, Gemini and Copilot is: to the provider’s cloud, with contractual guarantees. For most companies, that is enough. But not for everyone: medical data, attorney–client privilege, banking and insurance matters, regulated industries, or simply an owner who says the company’s data stays within the company’s walls. This part is for them.
What does an “open” model mean?
OpenAI does not let the model behind ChatGPT out of its hands: it runs only on OpenAI’s servers, and it can be reached only through the app or the programming interface. That is a closed model.
An open (more precisely, “open-weight”) model, by contrast, can be downloaded. The developer—Meta, Mistral, DeepSeek—releases the trained model itself, and anyone can run it on their own computer, their own server, their own cloud. This is the most important difference:
- The data does not leave. The question and the answer are produced on the company’s machine. There is no third party.
- No per-user fee. The model is free; the hardware and the operation cost money.
- No dependence. If the provider raises prices, shuts down or changes its terms, the downloaded model stays.
- Customizable. The model can be further trained on the company’s own material—that is expert work, but it is possible.
Open models are not lower quality by default; they typically run one step behind the top. What is available in closed models today will be in open models in six months to a year. For many tasks, that is more than enough.
Who are the main players?
Meta—Llama
Facebook’s parent company is the pioneer of open models. The Llama family comes in several sizes: the smaller ones run on a powerful laptop, the larger ones on a server. A great many products are built on it behind the scenes, and its developer community is the largest. As a consumer, you meet Meta AI in WhatsApp and Messenger, which is built on the same thing.
Mistral—the European answer
Paris-based Mistral is the largest European AI developer. It releases open models, but also offers its own assistant and a cloud service for businesses, on European servers, under European law. For a company that wants the convenience of the cloud, but within the EU, and does not want to depend on an American provider, this is the most obvious choice. In daily work, their assistant can be used much like ChatGPT: it writes, summarizes, searches, analyzes files, creates images.
DeepSeek and Qwen—the Chinese open models
In early 2025, DeepSeek shook up the market by openly releasing a model competitive with the top models, at a fraction of the development cost. Alibaba’s Qwen family is similarly strong and open. Downloaded and run locally, these too can be used safely, since the data goes nowhere. We do not, however, recommend the Chinese companies’ online services for European business use: the data processing is opaque, and the models answer with filtering on certain topics.
The big players’ open variants
Google, OpenAI and others also release smaller open models, partly so that developers stay within their family. These are typically made for smaller, fast tasks.
What does “running locally” mean in practice?
There are two levels, and the difference matters.
On an individual computer
On a current, more powerful laptop or desktop, a mid-sized open model runs comfortably. There are free programs for this, such as Ollama or LM Studio: you download one, pick a model, and a few minutes later you are chatting in a ChatGPT-like window—without internet, entirely locally. This is excellent for trying out what an open model can do, or for processing confidential material without uploading it anywhere. On a good machine, the speed and quality are surprisingly good, but they do not compete with the top models.
On a company server
The more serious level: a server with one or more powerful graphics cards, running a larger open model that all of the company’s staff can reach through an internal web interface—like your own internal ChatGPT. The data stays in the server room, the number of users does not raise the fee, and the model can also be connected to the company’s document repository. This is an IT project: hardware has to be bought or leased, operated, updated and made securely accessible. It is no more complicated than a file server or an email system, but no simpler either.
In a European cloud
The middle road: the model runs not at the company, but on the machines of a European provider—Mistral, for example, or an EU cloud provider. The data stays within the EU and a European contract covers it, but the hardware and the operations are not the company’s worry. For many Hungarian companies, this is the realistic balance.
What is all this good for at a Hungarian company?
An open model running locally or in a European cloud is not a “weaker ChatGPT” but a different kind of tool. It pays off where:
- The data is confidential. Processing, summarizing and searching material that counts as a medical, legal, financial, personal or business secret.
- The task is well defined. Classifying incoming mail, extracting summaries from documents, querying an internal knowledge base, pulling data out of forms. In these, open models perform on par with the top models.
- The volume is large. If thousands of documents a day have to be processed, a per-user or per-query fee quickly becomes more expensive than your own server.
- It is required. In regulated industries or at companies under audit, “nowhere” is the simplest answer to the question “where does the data go?”
And it does not pay off where top quality is needed for a hard, open-ended task—a complex analysis, a delicate customer text, an intricate programming job. For these, the closed top models are still better today, and the data protection guarantees of their business subscriptions are sufficient in most cases.
Where to be careful
- “Open” does not automatically mean “secure.” The downloaded model is safe, but the system built around it has to be protected like any other internal application: who has access, what they see, what is logged.
- Operations are real work. An open model does not update itself. Someone has to keep an eye on new versions, security flaws and performance.
- Quality is uneven. In Hungarian, some open models are weaker than in English. Do not plan on one without trying it.
- Hardware can be expensive. A server for a more serious model costs several million HUF. Leasing it, or running in a European cloud, is often cheaper in the first year.
How to try it
- On a laptop, in half an hour. Download LM Studio or Ollama, pick a mid-sized model, and give it the same task you usually give ChatGPT. The difference is instructive—in both directions.
- With a confidential document. One you would never upload anywhere. See what it can do with it locally.
- If you need the company level, request an assessment: for what task, in what volume, with what data—that is what reveals whether your own server, a European cloud or a good business subscription is the right answer.
Who are open and European models for?
For the company where the location of the data is not negotiable. For those who want to automate many repetitive, well-described tasks in large volumes. For those who do not want to depend on a single American provider. And for those who like to understand, and keep in their own hands, what runs at their company.
In the next part we move on to the more spectacular side of AI: creating images, voice and video—what it is good for, what it is not, and what to watch out for when you use it in the company’s name.
Do you work with data that cannot leave the company?
If you have a question, a request or a specific need related to this topic, write to us. We will look at which path—your own server, a European cloud or a business subscription—fits your data and your budget.
Triox Informatika Kft.