whose software is it, really?
Sovereignty comes down to two plain questions: where does your software run, and where does your data go? This page answers both, and it's honest about what's contractual today versus what we're still putting in writing.
where it runs. where your data goes.
Sovereignty is often confused with who trained a model's weights. That's the wrong question. What matters for your institution is the path your work travels: the servers that process your prompts, and the jurisdiction that governs them. Get those two right and you have a tool a policy office can actually approve.
An open model is a file, not a service. A closed API is a live connection to someone else's servers — the thing that can send your work abroad. An open model is the opposite: a file of numbers we run on French soil, with no company behind it and nowhere to phone home to. Its makers' location can't change the path above.
Your source and prompts cross the Atlantic. The service can be repriced, throttled, or switched off by a decision made an ocean away, and it can be reached by foreign subpoena. You can't see inside it, can't host it yourself, and can't sign off on it. The capability is rented — and rent can be cancelled.
The weights are downloaded, inspectable mathematics running on servers in France. Your prompts and code stay in the EU. No foreign company can flip a switch on it, because nothing in the path depends on one. This is sovereignty by architecture — not by a promise someone else can break.
Revocable by no one. That's the test. If a single foreign decision can take your tool away, you never really had it. We build so the answer to "can this be cut off?" is no.
eu-sovereign, all the way down.
MakeMode is a thin, transparent layer that turns proven open components into something an institution can own, inspect, and host itself. Nowhere in the path does your work leave Europe.
compute & inference
Served by Scaleway Generative APIs in France (fr-par). We run no GPUs of our own — one named EU processor, one standard GDPR DPA, France-only hosting.
identity & storage
EU identity and institutional SSO; your code, published work and research stored on European, GDPR-native infrastructure. Nothing routed through a US host.
the app itself
Open source, no telemetry, nothing phoning home. You can read every line, fork it, and self-host it — the platform can't lock you out of your own tool.
how we secure it → · what we do with your data → · how to take it with you →
eu-provenance by default. glm optional. all on french soil.
We default to models that are both EU-trained and EU-run, which is the clean procurement story. We also offer an open-weights model of Chinese origin, hosted in France, if you want it. Either way your prompts never leave the EU.See the full models page → for how they compare on speed and benchmarks.
mistral small 3.2 (Mistral, France) — the fast everyday model.
devstral 2 (Devstral, EU) — the agentic coding model.
mistral medium 3.5 (Mistral, France) — the higher-quality flagship, for detailed pages.
European weights and a European host. No foreign-origin footnote at all — the cleanest answer a procurement officer can ask for. Served by Scaleway in France; your data stays in the EU.
glm 5.2 — a strong open-weights reasoning model you can switch on if you want it.
Open weights are downloaded mathematics. The model runs on French GPUs through the same EU host as the others — your prompts, your code, your outputs never reach China or the US. Origin of the weights is not the path of your data.
The key point, said plainly: open weights are a file of numbers that runs wherever you put it. We put it on French soil. So with GLM you get a capable model whose weights came from China but whose processing happens entirely in the EU — nothing about your work travels east, or west. EU-provenance first; GLM if you choose it; all served by Scaleway in fr-par.
| Model | Best for | SWE-bench Verified |
|---|---|---|
| Mistral Small 3.2fastest · lighter drafts | quick pages, simple tools | general model |
| Devstral 2agentic coding, EU | coding-heavy builds | 72% |
| Mistral Medium 3.5EU flagship | detailed, polished pages | 78% |
| GLM 5.2open weights, EU-hosted | most polished output | top open-weights † |
† GLM 5.2 leads open-weight models on LMArena; its makers report 62% on the harder SWE-bench Pro (a different, tougher test — not comparable to Verified). Rankings shift weekly, so see the live boards: LMArena · Artificial Analysis · SWE-bench.
open is arguably more sovereign, not less.
It feels backwards, so here it is in full: a closed US API is a black box you rent and can't see inside. An open model is a file you hold. Which one is more truly yours?
inspectable
The weights are downloadable and the agent is open source. Your own staff, or an auditor, can look at what runs — not take a vendor's word for it.
forkable, no lock-in
No single company owns the path. Swap the model in one line, move hosts, or self-host on your own EU cluster. You're never bound to one lab or one country.
no remote kill switch
A closed API can be revoked by terms, by price, or by government order. A model running on your continent, on infrastructure you can name, can't be switched off from abroad.
The sovereign question was never "who made the math." It's "who controls where it runs, and who can take it away." On both, open weights on European soil win.
honest about what's architecture and what's signed.
We'd rather under-claim than mislead a policy office. Sovereignty here is real architecture plus an operating commitment that becomes contractual as data-processing agreements are signed. We publish what's true today and what we're still closing.
The app is open source, runs no telemetry, and targets an EU inference host. Inference is served by Scaleway Generative APIs in France; we operate no GPUs of our own. The default models are EU-provenance (Mistral, Devstral). Scaleway publishes France-only hosting, a standard GDPR DPA, and a no-collection / no-training stance on prompts.
EU-only residency and no-exit are our design intent and operating goal, becoming contractual as DPAs are signed — not, at this stage, asserted as settled guarantees. The GLM-on-EU-hosting path in particular is a credible, aspirational path: the weights are Chinese-provenance, the hosting is European, and the no-proxy and signed-DPA details are what we're closing in writing before we present any of it as a guarantee.
This honesty is the point. A sovereignty claim you can't audit is exactly the kind of black box we're built to replace. read the full trust overview →
build on ground you own.
EU-provenance models by default, open weights on French soil if you choose them, and nothing in the path that a company abroad can switch off. The rest of the trust story is here too.