A 100% local AI chat, hosted in France
Serge is a fully local AI chat interface, powered by LLaMA.cpp : no API key, no external calls. Installed and supervised by DINAO, hosted in France — your prompts never leave the server.
What is Serge?
Serge is a self-hosted AI chat interface and entirely local, built around LLaMA.cpp. With no API key or external service calls, it allows you to converse with open-weight language models directly from the browser, in a clean interface. Downloading and switching between GGUF / LLaMA models is done directly from the application.
Designed to be simple to deploy (a single Docker container combining the Python/FastAPI API, the frontend, and the inference engine), it is suitable for organizations that want a conversational AI with no data leakage. Quantized models run on CPU ; a GPU accelerates inference and allows for larger models.
DINAO hosts it in France, manages the models, updates, backups, and supervision — for truly sovereign AI.
Host Serge at DINAO
Resource tiers compatible with Serge prerequisites (minimum 2 vCPU / 8 Go / 10 Go). Hosted in France, fully managed.
- 4 dedicated vCPU
- 8 Go RAM
- 80 GB NVMe
- Daily backups
- Managed & monitored by DINAO
- 8 dedicated vCPU
- 16 Go RAM
- 160 GB NVMe
- Daily backups
- Managed & monitored by DINAO
2 hidden tier(s) (insufficient resources for this app) : Découverte, Standard
This application uses AI
The container hosts the application, not the AI engine (which requires dedicated GPUs) : inference is handled externally, with your own provider key. Prioritize a sovereign engine — Mistral AI, NumSpot, Scaleway, or OVHcloud AI Endpoints (France, GDPR) ; an international provider (OpenAI, Anthropic…) only if a specific capability requires it. Inference subscriptions are not included in hosting.
Technical details
You might be wondering…
Does Serge send my messages to OpenAI or elsewhere?
No, never. Serge performs inference locally with LLaMA.cpp. No API key, no outgoing calls: your prompts and responses remain entirely on your server in France.
What models can I use?
Open-weight models in GGUF format (LLaMA, Mistral, etc. families), downloadable and changeable from the interface. The recommended size depends on your plan and the presence of a GPU.
Do I need a GPU?
No to get started: small quantized models run on CPU. A GPU (optional from the Team plan) significantly accelerates performance and allows for larger models.
Is my conversation history private?
Yes. It is stored on your dedicated instance, in France, and is neither shared nor used to train anything.
What is the difference with a standard ChatGPT?
Serge prioritizes absolute confidentiality: everything is local. Quality depends on the chosen open-weight model, but no data leaves your infrastructure, unlike cloud services.
