French open-source AI: the companies, the models, and where they run
French open-source AI rests on a few companies with distinct jobs: Mistral AI publishes models, Hugging Face hosts the world's open weights, and OVHcloud and Scaleway run models in French data centres. The origin of a model and the jurisdiction of your data are separate questions: a Chinese or American open-weight model run by a French operator in France stays out of the CLOUD Act's direct reach, while a French-founded platform can be a US company. Evanesca works at the last step, running models from Mistral AI, the Swiss AI Initiative, Qwen, DeepSeek, Z.ai, OpenAI, Meta and Google in Europe, most of them in France.
Published 6 Oct 2026 · checked on 20 Oct 2026
Four layers, four kinds of company
“French AI” mixes companies that do very different things. Sort them by layer before asking which ones matter for your data:
| Layer | What it does | French example | Where the company answers |
|---|---|---|---|
| Models | Trains models and publishes the weights | Mistral AI | France |
| Hub | Stores and distributes the weights | Hugging Face (French founders) | United States |
| Operators | Run the weights on their GPUs and sell access | OVHcloud, Scaleway | France |
| Services | What you use: a chat, an API | Evanesca, among others | France |
For the jurisdiction of your prompts, the third layer decides: the company whose servers run the model. The maker's nationality and the hub's do not.
Mistral AI, the French model maker
Mistral AI was founded in April 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix (about), and it is a French company registered in Paris (legal notice). Its founders came from the two labs that shaped open models: Mensch co-wrote DeepMind's Chinchilla paper, and Lample and Lacroix co-wrote Meta's LLaMA paper.
“Open” does not mean the same licence for every release. Mistral Small 3.2, which Evanesca runs at Scaleway in Paris, is under Apache-2.0, which allows commercial use. Mistral Large 2, by contrast, came out in July 2024 under the Mistral Research License, which allows research and non-commercial use only; commercial use needs a separate licence (announcement). Read the licence of each model, not the reputation of its maker.
Hugging Face: French founders, US company
Hugging Face was founded in 2016, in New York, by three French entrepreneurs — Clément Delangue, Julien Chaumond and Thomas Wolf (source). Every open-weight model that Evanesca runs has its weights published there, so it is the first stop for anyone who downloads weights. The company itself is Hugging Face, Inc., “a Delaware corporation”, and its terms are governed by New York law (terms of service). On 3 September 2026, NVIDIA announced that it had agreed to acquire it, and said that Hugging Face would remain an open platform (NVIDIA).
The lesson is not about Hugging Face's quality: a French-founded platform is not European infrastructure. Downloading weights from the hub is harmless for your data; sending prompts to a service run by a US company is a different matter, whatever the founders' passports.
OVHcloud and Scaleway: where the models run
OVHcloud runs its AI Endpoints in Gravelines, in northern France. Its operating company, OVH SAS, belongs to OVH Groupe, controlled by the Klaba family (shareholders), and the service states “zero data retention: we keep only the data required for billing purposes” (AI Endpoints). For Evanesca it has been a fallback since 20 October 2026: it takes over Qwen3.8 27B, GPT-OSS 120B, Mistral Small 3.2 and Llama 3.3 70B when Scaleway, their first operator, fails.
Scaleway, a company of the iliad group, runs its Generative APIs in Paris (about Scaleway). It keeps no prompt by default, but may keep and read requests that cause errors or look abusive, for up to 14 days (data privacy).
Neither has a US parent, which puts the models they run out of the CLOUD Act's direct reach. Both groups have entities in the United States — OVHcloud's and Scaleway's — which is why “direct” matters: hosted in France is not the same as immune (why EU hosting is not enough). Evanesca's sovereignty page details the route, and the passport under each answer names the operator that ran it.
A model's origin is not your data's jurisdiction
Most of the open-weight models people use in France were trained elsewhere. What keeps them under European rules is the operator. Six examples from Evanesca's catalogue:
| Model | Maker (origin) | Size | Licence | Runs in |
|---|---|---|---|---|
| Mistral Small 3.2 | Mistral AI (France) | 24B | Apache-2.0 | Paris, Scaleway (fallback: OVHcloud) |
| Qwen3.5 397B | Qwen (China) | 397B (17B active) | Apache-2.0 | Paris, Scaleway |
| GLM-5.2 | Z.ai (China) | 753B | MIT | Paris, Scaleway |
| DeepSeek V4 Flash | DeepSeek (China) | 284B (13B active) | MIT | Paris, Scaleway |
| Llama 3.3 70B | Meta (United States) | 70B | Llama 3.3 Community License | Paris, Scaleway (fallback: OVHcloud) |
| GPT-OSS 120B | OpenAI (United States) | 117B (5.1B active) | Apache-2.0 | Paris, Scaleway (fallback: OVHcloud) |
The DeepSeek row shows the point best. DeepSeek's own app stores personal data in China, according to its privacy policy; the same family of weights, run by Scaleway in Paris, sends nothing there. DeepSeek V4 Flash, hosted in France explains the difference, and the full list gives every model's operator and place.
What European rules ask of open models
Open weights do not take a model out of the GDPR. The European Data Protection Board considers that a model trained on personal data cannot, in all cases, be considered anonymous: it has to be assessed case by case, by how likely personal data can be extracted from it (Opinion 28/2024, 17 December 2024).
Nor do they take a maker out of the AI Act. Since 2 August 2025, makers of general-purpose models must document them, inform the providers who build on them, follow a copyright policy and publish a summary of their training content (art. 53(1)). A free and open-source licence with public weights spares only the first two duties, and never for a model with systemic risk, presumed above 10²⁵ floating-point operations of training (art. 51(2) and 53(2), Regulation (EU) 2024/1689). Evanesca's AI Act page sets out who owes what.
What open weights do give you is choice: you can inspect the model card, test the model yourself and, above all, pick the operator — which is where jurisdiction is decided.
Three ways to use French open-source AI
- Through a service built mostly on French operators. Evanesca is one: a private AI agent and chat in the browser and an OpenAI-compatible API, with every model run in Europe and no copy of your prompts kept by Evanesca.
- Through the operators directly. OVHcloud AI Endpoints and Scaleway Generative APIs sell API access to open models; you contract with them and pick the models yourself.
- On your own servers. Inference engines such as vLLM, SGLang or Ollama run the weights on machines you control — ideally at a provider with no US control, since the EU region of a US cloud stays within the CLOUD Act's reach.
For the third route, self-hosted AI gateways compares the tools; for the first, the models and the API docs show what Evanesca runs. Evanesca is not open yet; this site shows how it will work.
Frequently asked questions
Which open-source AI models are French?
Those of Mistral AI, a French company registered in Paris. Mistral Small 3.2, under Apache-2.0, runs on Evanesca at Scaleway in France. Check each licence: Mistral Large 2, for instance, came out under the Mistral Research License, which allows research and non-commercial use only.
Is Hugging Face a French company?
It was founded in 2016, in New York, by three French entrepreneurs, but the company is Hugging Face, Inc., a Delaware corporation whose terms are governed by New York law. On 3 September 2026, NVIDIA announced an agreement to acquire it.
Does a Chinese model run in France send my data to China?
No, when a French operator runs the published weights in France. On Evanesca, the Chinese models run in Europe — at Scaleway in Paris or AKI.IO in Germany, with OVHcloud in Gravelines as a fallback for Qwen3.8 27B — and your prompts reach neither the model's maker nor China. The operator decides the route, not the maker.
Are open-source AI models GDPR-compliant?
Not by default. The European Data Protection Board considers that a model trained on personal data cannot, in all cases, be considered anonymous (Opinion 28/2024, 17 December 2024). Compliance depends on how and where the model is deployed: who processes your data, under which contract, and what is kept.
Read next
- DeepSeek V4.1 Flash: what it is, and where it runsDeepSeek V4.1 Flash: a 552B multimodal MoE with a 1M context, under MIT, released on 10 September 2026. On Evanesca, AKI.IO runs it in Germany.
- Tool calling on open-weight models: why it depends on who runs themTool calling turns a model into an agent, but it depends on the serving stack as much as on the weights. How to check it in one call, and which models pass.
- Open-weight reasoning models: what thinking costs, and how to choose oneReasoning models think before they answer, and that thinking is billed as output. Which open-weight models reason, what it costs, and how to choose one.