What is an LLM?
An LLM is a model trained on immense volumes of text in order to predict and generate language. It is the part that drafts, summarises, translates and answers questions. Its "size" is measured in parameters (for instance 24 billion, written 24B): the more it has, the more capable it is, but the more GPU memory and energy it consumes.
The point is not to have the biggest model, but the model best sized for the use case, and to keep control of it from end to end.
Open source ≠ open weight: the distinction that matters
Open weight
The model weights (the outcome of training) are published and downloadable. You can run the model on your own servers, adapt it, and operate it offline.
This is the minimum requirement for sovereignty: the model lives on your side. Examples: Mistral, Llama, Qwen, Gemma.
Open source (in the strict sense)
This goes further: alongside the weights, the training code, the data and the licence make it possible to reproduce, audit and fully modify the model.
Maximum transparency. Most "open" models are in reality open weight; fully open source models remain rarer.
Why this concerns you: a closed proprietary model (ChatGPT, Gemini, Claude through an API) gives you neither the weights nor control: your data leaves your perimeter and you depend on the vendor. An open weight or open source model runs on your infrastructure: that is the foundation of sovereign AI.
A stack agnostic of the model, which scales with the GPU
Our architecture does not depend on any one model in particular. The LLM is an interchangeable component: we can deploy the one that fits your need, your language and your budget, and replace it later without rebuilding the platform.
Interchangeable
The interface, the RAG layer and the monitoring stay identical; only the model changes. Zero dependency on a single vendor.
Scales with the GPU
The size of model you can run depends on GPU memory (VRAM). More resources = a larger model, or more simultaneous users.
Sized to the need
We pick the model / GPU pairing as precisely as possible: no costly oversizing, and no under-capacity.
A concrete example: a 24B model fits on a single A100 80 GB GPU and serves more than 200 simultaneous users. Need a larger model, or more throughput? We add GPU. The principle stays the same: see our technical stack.
The most sovereign LLMs
A selection of open models (open weight) suited to a sovereign deployment, with a focus on French and on Europe. The landscape moves fast: this list is indicative.
| Model | Vendor | Sovereign advantage |
|---|---|---|
| Mistral Small 4 | Mistral AI 🇫🇷 | French vendor, excellent in French, open weight, sized to run on a single GPU. The best sovereignty / performance ratio for Europe. |
| Mistral Small 3.2 24B | Mistral AI 🇫🇷 | The version running in production for our clients. Robust, proven, strong in French. |
| Llama 3.x | Meta 🇺🇸 | Open weight, very large ecosystem. Non-European vendor, but fully local execution is possible. |
| Qwen 3 | Alibaba 🇨🇳 | Strong multilingual performance, open weight. To be assessed against the sensitivity of the context. |
| Gemma | Google 🇺🇸 | Efficient compact models, open weight, executable locally. |
Maximum sovereignty = European vendor + local execution. Mistral ticks both boxes, which makes it our default choice for French public bodies.
Which model for your organisation?
We help you choose the LLM best suited to your language, your use cases and your GPU budget, then we deploy it on sovereign infrastructure.