
Sovereign AI for SMEs: what Mistral Large 4 and Ecosia's departure change

Mistral launches Large 4, Ecosia leaves Mistral: what this debate changes for SMEs, and how to choose a sovereign, GDPR-compliant AI without lock-in.
In a single week, European AI saw two announcements that almost contradict each other. On 6 October, Mistral AI presented Mistral Large 4, its most powerful model to date, trained in Europe and soon to be available with open weights. The same day, the German search engine Ecosia told POLITICO that it was leaving Mistral, « déçu par la qualité » (“disappointed with the quality”), in favour of open models, several of them Chinese.
For an SME manager, this debate may seem remote. Yet it raises a very concrete question. If you bring AI into your tools, who are you entrusting your data to? And what happens the day your provider no longer suits you? Here is a jargon-free overview to help you choose a sovereign AI suited to your SME.
What happened: Mistral Large 4 and Ecosia's departure
Mistral Large 4, the European “Chonk”
On 6 October, Mistral opened a public preview of Mistral Large 4, nicknamed “the Chonk”. According to the official announcement, it is a multimodal model with 1,000 billion parameters, trained from scratch in Mistral's own data centres in Europe. The company reports very good results in coding, agentic workflows and cybersecurity, and plans to publish the model's weights by the end of October.
Two points directly concern businesses. First, Mistral plans a European deployment “operated end to end” by itself, under European law. Second, once the weights are published, the model can be installed on a private cloud or on in-house servers, which lets you keep control of your data.
Ecosia: disappointed with the quality
Ecosia had replaced OpenAI with Mistral in May 2026, a choice presented as a gesture of European independence. Five months later, its founder Christian Kroll is backtracking. He speaks of models « un an en retard » (“a year behind”) the competition and of regularly overloaded servers: « Nous étions tout simplement un client trop important pour Mistral » (“we were simply too big a customer for Mistral”), reports BFMTV.
Ecosia now goes through Melious, a European platform that hosts several open models such as Qwen, GLM and Kimi. According to Christian Kroll, this change cut costs roughly in half while improving quality. He later told EU-Startups that his choice was about open models, not about China: the models run on European infrastructure and the data is still processed in Europe. For its part, Mistral invited Ecosia to test Large 4 as soon as possible.
Open-weight, sovereignty: what it means in practice
Both stories rest on notions that are often confused. Let's sort them out.
- A closed model (ChatGPT, Claude, Gemini) can only be used through its publisher's service. You send your data, you receive an answer, and you depend entirely on its terms, its prices and the location of its servers.
- An open-weight model is published so that anyone can download it and run it on the infrastructure of their choice. This is what Mistral plans for Large 4, and what Ecosia uses through Melious.
- Sovereignty, finally, does not depend only on the nationality of the publisher. It depends above all on where the servers run, which law applies to them, and your ability to change provider.
That is the whole lesson of the Ecosia case: a model designed outside Europe can be run on European servers, and a European model can disappoint on quality or availability. Asking whether the AI is French is therefore not enough. The question to ask is rather: where does my data go, and can I change my mind tomorrow?
How to choose an AI model for your SME: 5 criteria
1. Your data, the GDPR and the Cloud Act
As soon as the AI processes customer data, contracts or HR information, the GDPR applies. Before choosing a model, ask three questions: where are the servers hosted, is the data you send used to train the model, and which law applies to the provider? A provider subject to US law may be affected by the Cloud Act, even when its servers are located in Europe. For sensitive data, European hosting operated by a European company, or an open model installed on your own servers, significantly reduces this risk.
2. Costs, today and tomorrow
Models are generally billed by the volume of text processed. For occasional use, the difference is small. For a tool used every day by a whole team, it quickly shows on the bill. Ecosia says it roughly halved its costs by changing model: this item can vary a great deal. Think in terms of cost per real use (one quote analysed, one email processed) rather than the advertised rate.
3. Dependence on a provider
AI models are evolving very fast. The best one today may no longer be the best next year, and a provider can change its prices or terms, or withdraw a model. If your business tool is built around a single provider, every change becomes a project. This is exactly what Ecosia had to deal with twice in five months.
4. Quality on your use cases
The benchmarks published by vendors show a trend, not a guarantee. What matters is the quality on your tasks: summarising your meeting notes, extracting information from your purchase orders, answering your customers in your tone of voice. The most reliable approach is still to test two or three models on a sample of your own documents before deciding.
5. Hosting and availability
An excellent model that is unavailable at peak times is useless. Check the provider's availability commitments, and plan a fallback solution. For critical uses, an open model hosted on infrastructure you control offers a stability that consumer services do not always guarantee.
Why design your tools so that you can change model
If there is only one thing to remember from this week: do not tie your business software to a single AI model. A well-designed integration separates your application (your screens, your business rules, your data) from the model that generates the answers. In practice, this means:
- a single integration layer, which lets you switch from one model to another through simple configuration;
- sensitive data filtered or anonymised before anything is sent to an external service;
- tests on your own use cases, to compare models objectively with each new version;
- the option of hosting an open model on your servers when confidentiality requires it.
This is how we integrate AI into the custom software we develop at METEORA. Our projects are hosted on our own infrastructure, in Switzerland and France, deployed with Coolify: we explain this choice in our dedicated article. And we use the best models of the moment ourselves for development, as we described with Claude Opus 5, without ever locking our clients into a provider.
The best AI model for your SME is the one you can replace without rebuilding everything.
And no, this approach does not necessarily cost more. Planned from the design stage, it represents a small share of a project's budget, and it avoids costly rebuilds later. For orders of magnitude, our article on the cost of custom software for an SME gives concrete price ranges.
In summary
Mistral Large 4 shows that Europe can produce first-rate models. Ecosia's departure, for its part, is a reminder that sovereignty cannot simply be declared: it is built into the architecture of your tools. For an SME, this means data hosted within a legal framework you control, costs assessed on your real uses, and the freedom to change model as the market evolves. This is exactly what we put in place at METEORA, in Rumilly near Annecy, for SMEs in Haute-Savoie and beyond that want to bring AI into their business tools.
To put this into practice, see our AI solutions for businesses.
Frequently asked questions
What is sovereign AI for an SME?
It is AI whose hosting, legal framework and provider dependence you control. In practice, your data is processed on servers located in Europe, subject to European law, and your tool can change model without being rebuilt. The nationality of the model's publisher matters less than where it runs and how it is integrated.
Is Mistral AI suitable for a business like mine?
Yes, Mistral AI is a serious option for a business that wants European AI. Its new Large 4 model targets professional uses and will be available with open weights, so it can be installed on private infrastructure. As with any model, the safest approach is to test it on your own use cases and keep the option of changing.
Is using ChatGPT or another AI compatible with the GDPR?
It can be, provided the use is properly framed. You need to know where the data is processed, make sure it is not used to train the model, sign the appropriate contractual commitments and avoid sending unnecessary personal data. For sensitive data, European hosting or an open model on your own servers is easier to justify.
Is an open-weight model necessarily safer?
No, not automatically. An open-weight model gives you the option of hosting it yourself and therefore keeping control of your data, but security then depends on the infrastructure, access management and maintenance. Well hosted and well integrated, however, it is one of the most controllable options for an SME.
How can you avoid depending on a single AI provider?
By separating your business software from the AI model it uses. A single integration layer lets you switch from one model to another through configuration, and tests on your own data let you compare models with each new version. This architecture is planned from the design stage of the project.
How much would your project cost?
Project type, features, timeline: a costed estimate in two minutes, with no commitment.
Want to bring AI into your tools without losing control of your data?
From Rumilly, near Annecy, we help you choose the right model and integrate it into your business software, hosted in France or Switzerland.