Back to the blog

EU-hosted LLM providers in practice: Mistral, Scaleway, IONOS and EU GPT

· By Bee Flow · 5 min read

Four ways to run inference in the EU without buying GPUs. What each provider is, how our adapter talks to it, and the differences that show up in the code.

Four lighthouses of different heights in a row, one with its light on

Some organisations want a hosted model and also want inference to happen inside the EU. Next to the self-hosted runtimes, Bee Flow supports four providers for that: Mistral through its EU endpoint, Scaleway Generative APIs, the IONOS AI Model Hub and EU GPT. Each is one file in server/core/providers/, created through a single factory, and the orchestration code never branches on a provider name. An adapter holds only what is different about its provider, which makes the differences easy to read.

This post walks through them. It has no prices and no benchmarks. Model lists change often, and all four are read live from the provider's API.

At a glance

Mistral Scaleway IONOS EU GPT
Protocol Mistral SDK v3 OpenAI chat completions OpenAI chat completions POST /v1/responses
Region api.eu.mistral.ai Paris, Amsterdam Berlin Scaleway EU
Who picks the model You You You Their router
Your tools in agents Yes Yes Per model No
Embeddings Yes Yes Yes No
Token usage reported Yes Yes Yes No

Of the four, only IONOS has a rerank model in our catalog.

Mistral

The adapter uses the official @mistralai/mistralai SDK, version 3. That SDK strips every request key it does not recognise, without an error. Parameters therefore use the SDK's own camelCase names, and mistral.test.js runs them through the SDK's real outbound schema, so a dropped key fails a test.

A stored base URL is honoured only when it is a Mistral host: api.mistral.ai or the regional api.eu and api.us endpoints. Anything else is ignored with a warning, so a typo cannot send traffic to a third party. For processing in the EU you set the provider's address to https://api.eu.mistral.ai.

Mistral accepts tool-call ids of exactly nine letters or digits and answers anything else with a 400. An id from another provider, left over from a model switch halfway through a conversation, does not have that shape. We hash those ids deterministically, so a call and its result keep matching and the replayed prompt prefix stays byte-identical, which Mistral's prompt cache depends on. Only a conversation key goes out as prompt_cache_key, never a user id, so that field carries no personal data.

Reasoning is a switch on Mistral Small 4 and Medium 3.5, with none and high as the documented values, and Bee Flow maps its own effort levels onto those two.

Scaleway Generative APIs

Serverless, OpenAI-compatible endpoints for open-weight models. The default address is https://api.scaleway.ai/v1, and https://api.scaleway.ai/<project-id>/v1 scopes calls to one Scaleway Project. On a self-hosted server you can set both from the environment:

SCALEWAY_API_KEY=<your-secret-key>
SCALEWAY_URL=https://api.scaleway.ai/<project-id>

The adapter adds three things to the generic OpenAI path. Reasoning is on by default at Scaleway, and the accepted reasoning_effort values differ per model: gpt-oss-120b has no none, GLM speaks high and max. An unsupported value is a 400 that fails the whole call, so the adapter drops a value the model does not know. It clamps max_tokens to the model's serverless output cap, because asking for more is rejected outright. And it sets stream_options.include_usage, without which a streamed turn would report zero tokens.

Cost accounting is the subtle part. Every Scaleway model is open-weight and also sold by other hosts, so a fuzzy name match would price a call at some other host's rate. Bee Flow prices Scaleway usage only through the exact scaleway/<vendor>/<id> key. The same secret key also enables Scaleway's Whisper transcription.

IONOS AI Model Hub

Also OpenAI-compatible, hosted in Germany, with one endpoint: https://openai.inference.de-txl.ionos.com/v1. IONOS is the one provider in this group that covers chat, embeddings and reranking, so a knowledge base can use a single vendor for all three. For embeddings use BAAI/bge-m3 (1024 dimensions). The Qwen3-VL embedding model returns more dimensions than our vector index accepts (2000), so it cannot back a knowledge base.

Models are matched by a normalised name, so a suffix such as -Instruct or -fp8 does not matter. A model that IONOS serves and our catalog does not know still works, but it carries no price, because Bee Flow never guesses one. reasoning_effort is sent only to models with a known vocabulary (gpt-oss) and dropped for the rest.

EU GPT

EU GPT serves open-weight models from Scaleway's EU regions behind a single endpoint shaped like OpenAI's Responses API. Most of the adapter deals with how narrow that endpoint is. There is no /v1/chat/completions. The router always picks the model: the value on the wire is auto, which we list as eugpt-auto, and the model it chose comes back on the response.

There are no client tools either. EU GPT runs four of its own on its side (web search, web fetch, a calculator, the current date) and reports them in the stream. We never forward those as tool calls, or our tool loop would try to run something the provider already ran. For agents that need your own tools, EU GPT is the wrong choice.

EU GPT can keep earlier turns as server-side state through a conversation_id, which would put your history in their store. We send the history as a transcript inside the user message instead. It goes in the user turn and never in instructions, so quoted user text cannot gain the authority of the system prompt. Images are dropped, and since the response carries no token usage, a turn reports none instead of an invented count.

What stays the same

All four appear in the same model tier picker, and the Privacy Shield scan runs on your Bee Flow server before a message goes to any of them. Where inference runs is one part of a GDPR assessment. Your agreement with the provider is another part, and no adapter settles that for you.

Setup guides are in the docs for Scaleway and IONOS. The platform page shows where providers fit in, privacy and security covers what leaves your server, and self-host or EU cloud shows where Bee Flow itself can run.

To find out which of the four fits your organisation, plan a 20-minute call.

#ai-models #privacy