European AI for language models: a catalogue of EU-hosted models, a policy-aware router and an OpenAI-compatible API. One catalogue. One API. One jurisdiction.
Hosting entries describe configured endpoints, not proof of live GPU capacity. Mock inference runs no model weights and reports dev-mock. Check the deployment status before use. Status →
Every model Europe runs, in one place: model cards, providers, hosting regions, licenses, benchmarks and the documents behind them. Cards we have not checked are marked unverified.
One OpenAI-compatible endpoint chooses a model under your policy. Residency, budget and allow lists are decided before the request leaves, and the decision is written to the trace.
First-party inference on infrastructure LLM EU operates or tightly controls, in EU regions, under EU law. Open weights can be hosted on request.
Devstral Small 2 is a 24B Apache-2.0 model fine-tuned for software-engineering agents, built with All Hands AI. The model card reports 68.0% on SWE-bench Verified and 55.7% on SWE-…
EuroLLM-1.7B-Instruct is the smallest model of the EuroLLM family, trained on 4 trillion tokens across the same 35 languages as the larger sizes. It is Apache-2.0 licensed and its …
EuroLLM-22B-Instruct (2512) is the largest released EuroLLM model, covering the same European language set as the 9B version. The model card states that long-context data sources w…
EuroLLM-9B-Instruct (2512) is the long-context revision of the EuroLLM 9B instruct model, trained for the 24 official EU languages plus Catalan, Galician, Turkish, Arabic, Chinese,…
Mistral Small 4 is an Apache-2.0 mixture-of-experts model with 119B total parameters and 6.5B activated per token. It accepts text and images and generates text over a 256k-token c…
Whisper large-v3 is OpenAI’s multilingual speech-recognition and speech-translation model with 1.54B parameters under the Apache-2.0 licence. It has a fixed 30-second receptive fie…
Every endpoint carries a region and a hosting class. If your policy allows only the EU and no EU endpoint can serve the request, the request fails. It is not quietly sent somewhere else.
| Hosting classes | Sovereignty class |
|---|---|
| de-ber Berlin, Germany — LLM EU operated |
EU-owned, EU-operated
Company and infrastructure both under EU law and EU ownership. |
| de-fra Frankfurt, Germany — LLM EU operated |
EU-owned, EU-operated
Company and infrastructure both under EU law and EU ownership. |
| fr-par Paris, France — EU operator |
EU-owned, EU-operated
Company and infrastructure both under EU law and EU ownership. |
Describe what you need — transcription, German legal text, code review, embeddings — and the router maps the task to models that are documented for it in the regions you allow.
A verified card shows where the claim comes from: provider documentation or a license text. Where we have no source, the card says so. Verified is a documentation status, not a legal certificate.
The DPA is an unsigned template. Legal details and the subprocessor inventory are incomplete. No independent penetration test has been carried out.
Documents & current limitations →Providers correct their own cards. Researchers send evaluations we can cite. If a card is wrong, the correction is public and dated.
Put two to four models side by side on the fields that decide a deployment: context, price, region, license, languages, hosting, and what is verified.
Compare →A verified claim on LLM EU says where it comes from: a document you can open, or a value recorded on the endpoint row.
Benchmarks →