Sovereign AI: Swiss or European APIs, assistants, models and image generation
Infomaniak AI Tools, Mistral, Scaleway, OVHcloud, IONOS, Apertus, Euria, Lumo, Le Chat, FLUX: the Swiss and European equivalents of AWS Bedrock, ChatGPT and image generators, and how to make an AI solution reversible.
Published on 17 September 2026 by Lumineth
This is the family of tools that moves fastest, and the one where the most sensitive data sometimes flows without anyone noticing: a contract pasted into ChatGPT, a file summarised by an assistant. Fifth part of the digital sovereignty series, and the closest to what I put in place every day.
AI APIs: replacing AWS Bedrock
| Solution | Country | Origin | Models | Self-hostable | Pricing model | Note |
|---|---|---|---|---|---|---|
| Infomaniak AI Tools | Switzerland | Switzerland | Open source | No | Pay as you go | OpenAI-compatible API |
| Mistral La Plateforme | France | EU | Partial (open-weight) | No | Pay as you go | |
| Scaleway Generative APIs | France | EU | Open source | No | Pay as you go | |
| OVHcloud AI Endpoints | France | EU | Open source | No | Pay as you go | |
| IONOS AI Model Hub | Germany | EU | Open source | No | Pay as you go | |
| vLLM / Ollama (self-hosted) | Projects of American origin | International (open source) | Your choice | Yes | Free (GPU cost on top) | To run on European GPUs |
All these APIs expose an OpenAI-compatible interface: an application written for one works with the others by changing an address and a key. That is what makes the migration realistic. Infomaniak AI Tools is the Swiss option; Mistral brings its own models, among the best in Europe; Scaleway, OVHcloud and IONOS serve open models on GPUs in France and Germany.
Assistants: replacing ChatGPT, Claude or Gemini
| Solution | Country | Origin | Open source | Self-hostable | Pricing model | Note |
|---|---|---|---|---|---|---|
| Euria (Infomaniak) | Switzerland | Switzerland | Open-source models | No | Included with my kSuite | |
| Lumo (Proton) | Switzerland | Switzerland | Partial | No | Freemium | No history kept server-side |
| Le Chat (Mistral) | France | EU | No | No | Freemium | |
| Open WebUI / LibreChat | Community projects | International (open source) | Yes | Yes | Free | Interface to plug into a sovereign API |
For a team, the most controlled option is a free interface (Open WebUI or LibreChat) plugged into a Swiss or European API: conversations stay with you, the model changes with one setting, and you can connect your own documents.
Open language models
| Solution | Country | Origin | Open source | Self-hostable | Pricing model |
|---|---|---|---|---|---|
| Apertus (EPFL / ETH Zurich) | Switzerland | Switzerland | Yes (open-weight) | Yes | Free |
| Mistral open-weight models | France | EU | Yes (open-weight) | Yes | Free |
Apertus is the Swiss model, trained by EPFL and ETH Zurich with open weights. Mistral’s open-weight models are the European reference. Both run on your own hardware or through the APIs above. For the classification, extraction and summarisation tasks that make up most business cases, they are enough; for complex reasoning, American models keep a lead, which is then reserved for non-sensitive data.
Image generation: replacing Nano Banana
| Solution | Country | Origin | Open source | Self-hostable | Pricing model | Note |
|---|---|---|---|---|---|---|
| Infomaniak AI Tools (images) | Switzerland | Switzerland | Open-source models | No | Pay as you go | |
| FLUX / FLUX Kontext (Black Forest Labs) | Germany | EU | Partial | Yes (open-weight versions) | Pay as you go | Check the licence: some versions are non-commercial |
| ComfyUI + FLUX | American publisher | International (open source) | Yes | Yes | Free (GPU cost on top) |
The method: the model is a component
In a processing chain, the model should be just one block behind a gateway: the application talks to the gateway, the gateway talks to the current provider. You can then route sensitive data to a Swiss API and the rest to the best-performing model, and change your mind without rewriting. The next part details these technical tools.
I put this kind of migration in place with your teams: inventory, choices by data sensitivity, staged switch-over, operations. A first conversation is enough to know where to start.
Get in touchFAQ
Frequently asked questions
Are European models as good as American ones?
For classification, extraction, summarisation and everyday writing, yes. For complex reasoning, American models keep a lead in 2026. You choose by task and by data sensitivity.
What is Apertus?
A Swiss open-weight language model, trained by EPFL and ETH Zurich. It can be hosted on your own hardware or served by a Swiss API.
How do we avoid depending on an AI provider?
By placing a gateway between the application and the provider, using OpenAI-compatible APIs, and keeping your rules, prompts and evaluation data outside the provider.
A concrete case on your side?
Tell me what costs you time or what you want to take back control of; I get back within 24 business hours with a first read, no obligation.