For healthcare, finance and insurance, and the legal professions, our AI agents combine sovereign hosting, an anonymisation filter before any external call, and priced compliance deliverables (GDPR, AI Act, HDS/SecNumCloud). Human oversight remains the rule.
Document support, reports, patient reception — without ever exposing health data.
Document analysis, compliance and back-office support, under human control and fully traceable.
Assisted research, summarising and drafting, in strict respect of professional secrecy.
We reconcile external models (tokens at real cost) with data sovereignty: direct identifiers are pseudonymised before any call, then the answer is re-associated locally.
Documents, identifiers and sensitive data are processed within your own perimeter (local inference or an isolated resource in France).
Before any call to an external model, a filter pseudonymises, masks or removes sensitive data and direct identifiers.
Only what is strictly necessary, anonymised, is transmitted. The answer is re-associated with your context locally.
This arrangement makes our sovereignty claim credible and defensible for sensitive sectors (health, finance, legal), while keeping the transparency of the “tokens at real cost” model.
Try the anonymisation filter live →Priced compliance deliverables, optional, for regulated sectors — alongside the agent.
Prices exclude VAT. For multi-site or regulated projects, the in-depth audit fee is folded into the setup (terms of sale, art. 4 bis).
Our default, and our reason for being: a sovereign model running on your own premises. That is what we install unless you ask for something else. But the choice is yours — all nine models below are available, each with its origin, its size where published, and where it runs, that is, whether or not your data leaves your premises.
Frugality : most favourable — 24 billion parameters · runs on your own hardware
Artefact: mistralai/Mistral-Small-3.2-24B-Instruct-2506, Apache 2.0 licence. INSTRUCTION-TUNED, with tool calling. This is the default for our sovereign deployments.
Frugality : most favourable — 7 billion parameters · runs on your own hardware
Artefact: OpenLLM-France/Lucie-7B-Instruct-v1.1, Apache 2.0 licence (CNRS / Linagora). Its own model card calls the instruction tuning “light” and recommends retrieval-augmented generation: available on quotation, after fine-tuning and acceptance testing on your own documents.
Frugality : most favourable — 7 billion parameters · runs on your own hardware
Artefact: openGPT-X/Teuken-7B-instruct-research-v0.4 (Fraunhofer / OpenGPT-X). A RESEARCH build under a non-standard licence, with instruction tuning limited to English and German: available on quotation, after a licence review for your use.
Frugality : most favourable — 9 billion parameters · runs on your own hardware
Artefact: utter-project/EuroLLM-9B-Instruct, Apache 2.0 licence, funded by the European Union. INSTRUCTION-TUNED, 35 languages including the 24 official EU languages. Chosen when multilingual coverage matters most.
Frugality : cannot be determined — size not published by the vendor · hosted in France or in Europe
Large French model, hostable in Europe. Its size is not published.
Frugality : cannot be determined — size not published by the vendor · runs on your own hardware
Non-European publisher, but open weights: the model runs on your own hardware, with no outbound call. Several sizes, stated in the name of the version chosen.
Frugality : cannot be determined — size not published by the vendor · called at its publisher
Market model, called at its publisher. Size not published.
Frugality : cannot be determined — size not published by the vendor · called at its publisher
Market model, called at its publisher. Size not published.
Frugality : cannot be determined — size not published by the vendor · called at its publisher
Market model, called at its publisher. Size not published.
How to read the "frugality" column
It is an order of magnitude based solely on the parameter count published by the vendor, not a measurement of consumption. A model twice as large does not consume exactly twice as much, and we do not yet measure energy per request. We would rather say so than put forward a figure we could not defend to a buyer.
The most frugal model is not always the right choice. A model too small for the task fails, and the work has to be redone: three attempts on a small model consume more than one correct answer first time on a larger one. Frugality is judged on the work actually completed, never on the size displayed.
Models marked "Outside the EU" remain available to you: the choice is yours, and we would rather inform it than restrict it. Two distinctions matter. An open-weight model from a non-European vendor runs on your own hardware, with no data leaving it. A model called at its publisher, by contrast, means a transfer, and comes under our prior filtering and anonymisation options.
Table reviewed on 08/08/2026. Models and their sizes change: we keep this table up to date with every update of the site.
Available at order time (or through the audit, for regulated sectors).
Four questions to situate your need and get an order of magnitude. The exact price is confirmed by quotation.
Indicative, non-contractual estimate. Hosting with a provider qualified SecNumCloud by ANSSI, the French national cybersecurity agency; the choice of host remains adjustable according to your scope.
Book an audit: we scope the perimeter, the hosting (HDS/SecNumCloud) and the compliance deliverables that fit.
