- The best European models are on a par with mid-range American models, and sometimes ahead on French and European languages.
- They still lag on long reasoning and complex agents, where frontier American models keep the edge.
- Real sovereignty depends on where data is hosted and on reversibility, more than on the vendor's flag.
- For healthcare, the public sector, finance and personal data, a European model hosted in Europe is today a rational choice, not a militant one.
What we mean by sovereign
The word covers three different realities. Where the data is processed: an American model hosted in a European data centre by a European provider already answers part of the question. Who controls the model: an open-weights model you host yourself depends on no foreign commercial decision. And what it costs to leave: a system built on standard interfaces migrates in weeks, a system locked to proprietary functions does not migrate at all.
A company rarely needs all three. It needs to know which ones matter for it, and the answer differs for a hospital, a bank or a communications agency.
What the models are worth, on our real cases
We replayed our client test sets (field extraction from contracts, support reply drafting, request classification, meeting summaries, a six-step document processing agent) on the main European models available this summer, comparing them with a frontier American model and a mid-range American model.
On bounded tasks in French, the gap with the best American model is within the margin of error, and European models sometimes do better on French administrative and legal phrasing. On long reasoning and agent orchestration, the gap exists and shows: more failed steps, more retries. On the African languages we test for our clients in the CFA franc zone (Wolof, Bambara, Mooré), no model is truly good, but open models can be fine-tuned, which is a real advantage.
When to choose a European model
When data cannot leave
Healthcare, public sector, regulated finance, HR data: as soon as transfer outside Europe is a problem, a European model, or an open model hosted in Europe, becomes the obvious choice. Quality is sufficient for most use cases, and the legal risk disappears.
When the task is bounded and the volume high
Classification, extraction, large-scale summarisation: lower cost and equivalent quality tip the balance, sovereignty or not.
When reversibility is a criterion
Open-weights models, European or not, guarantee that you can always run your system, even if the vendor changes its prices, terms or owner. For a critical system, that guarantee is worth a few points of performance.
A European model called through an American platform, with data transiting through servers outside Europe, is no more sovereign than any other. Check the full path of the data, not only the model's name.
When not to choose it
For a complex agent that plans, calls tools and corrects its errors, frontier American models remain ahead, and the gap is paid in reliability. For code generation on large codebases, same finding. In those cases, the routing architecture remains the best answer: European or open model for the regular flow, frontier model for the steps that require it, with data anonymised before it leaves.
"Sovereignty is not a slogan, it is a property of your architecture. It is measured by looking at where the data goes and what it would cost you to change vendor."
Conclusion
European models have become a serious option, not a compromise. They win on French, cost and data control; they still lose on complex reasoning. The right decision is not made on a flag but on a test: your data, your tasks, your full cost. That test is what we now run systematically before any model choice.
FAQ
Is an American open model hosted in Europe sovereign?
On the data criterion, yes, if the host is European and nothing transits elsewhere. On the control criterion, yes as well since you hold the weights. What remains is dependence on the original ecosystem for updates.
Are European models AI Act compliant by default?
No. Compliance depends on your use, not on the model's origin. A European vendor makes documentation easier to obtain, which helps, but your deployer obligations stay the same.
How to test without committing a budget?
With a test set of two hundred examples drawn from your data, replayed on three models through their APIs. A week of work, a few dozen euros of compute, and a grounded answer.
What about African languages?
No model is yet satisfactory without adaptation. Open models allow fine-tuning on your own data, which we do for clients in the CFA franc zone with usable results on bounded tasks.
