EU-hosted AI models in 2026, model by model

What "EU-hosted" means, and what it does not

An EU-hosted AI model is a model whose inference — the computation that turns your prompt into an answer — runs on servers in the European Union, or in a country covered by an EU adequacy decision such as Switzerland. It says where. It does not say who: the same model can be EU-hosted by a French company or by a US company with a European region, and the legal picture is different.

That is why this page uses two labels, the same ones as our catalogue:

• EU sovereign — EU-hosted and run by a company under European control, such as OVHcloud, Scaleway, Mistral AI, AKI.IO or Infomaniak. These providers are outside the direct reach of the US CLOUD Act. That is not immunity from every foreign law: OVHcloud and Scaleway, for instance, each have a US entity.
• Fast access — served by US companies. Sometimes in an EU region, often not. Useful to try a new model early, never presented as sovereign.

Open weights are what make the first label possible for frontier-scale models: when a lab publishes its weights, any European provider with the hardware can serve them. Our article What is frontier AI? explains why, and What is sovereign AI? separates sovereign access from sovereign infrastructure.

EU-hosted models at a glance

The table below is generated from our live catalogue each time the page is built: a model that changes status changes column, and paused models are left out. Each family name jumps to its section, which also lists the other European providers serving it.

Today 17 models are live in the EU-sovereign lane on Frontière AI.

Model by model: who hosts what in Europe

For each family: where the publisher's own API runs, which versions you can call EU-hosted through Frontière AI and from which provider, and which other European providers listed it when we checked their price pages on 2026-10-13. Provider details — company, data centres, retention — are in European inference providers in 2026.

DeepSeek

DeepSeek publishes its weights under the MIT licence; its own API runs in China. On Frontière AI, EU sovereign: DeepSeek V4 Flash 0731 — Scaleway (France) and DeepSeek V4.1 Flash — AKI.IO (Germany), fallback Scaleway (France). Fast access only, served by US companies and labeled as such: DeepSeek V4 Pro. Other European hosts listed on 2026-10-13: DeepSeek V4.1 Flash at Scaleway (Paris), AKI.IO (Germany) and TensorX (Ireland, Finland); DeepSeek V4 Flash 0731 at Scaleway and TensorX. More on our DeepSeek page.

GLM (Z.ai)

Z.ai publishes GLM weights; its own API is run by a Singapore company and processes data mainly in Singapore. On Frontière AI, EU sovereign: GLM-5.2 — Scaleway (France) and GLM-5.3 — AKI.IO (Germany), fallback Scaleway (France). Fast access only, served by US companies and labeled as such: GLM-5.3 Flash. Other European hosts listed on 2026-10-13: GLM-5.3 at Scaleway (Paris), AKI.IO (Germany) and TensorX, and on Mistral AI's EU endpoint; GLM-5.2 at Scaleway, T-Systems (Germany), Regolo.ai (Italy) and TensorX. More on our GLM page.

Qwen (Alibaba)

Alibaba's Qwen team publishes most Qwen3-generation models under Apache-2.0, with a few under Qwen's own licences; check each model page. On Frontière AI, EU sovereign: Qwen3 235B — Scaleway (France), Qwen3.5 397B — Scaleway (France), Qwen3.6 35B A3B — Scaleway (France), Qwen3 Coder 30B — Scaleway (France) and Qwen3.8 27B — Scaleway (France), fallback OVHcloud (France). Fast access only, served by US companies and labeled as such: Qwen3.8 Flash. Qwen is the most widely hosted family in Europe: on 2026-10-13 it was also listed by STACKIT and T-Systems in Germany and by Infomaniak in Switzerland (Qwen3.5 397B), on top of OVHcloud and Scaleway. More on our Qwen page.

Mistral AI

Mistral AI is a Paris company. It publishes some models as open weights and sells others through its API. On its default endpoint, Mistral does not commit to an inference location; its EU endpoint runs in EU and EFTA data centres. It keeps API inputs and outputs for 30 rolling days for abuse monitoring unless zero retention is granted. On Frontière AI, EU sovereign: Pixtral 12B — Scaleway (France), Mistral Small 3.2 24B — Scaleway (France), fallback OVHcloud (France), Mistral Large 4 — Mistral AI (France) and Mistral Small 4 119B — Mistral AI (France). Mistral Large 4 is in "Public Preview" at Mistral: under § 4.3 of its terms, data sent to preview models may be used for training, and opt-outs do not apply. Mistral's open models are also served by OVHcloud and Scaleway, and Infomaniak listed Mistral Small 4 on 2026-10-13. More on our Mistral page.

Meta (Llama, Muse Spark)

Meta's Llama 3.3 70B ships under the Llama 3.3 Community License, which allows commercial use with conditions. Meta's newer Muse Spark models are proprietary and API-only: no European host can run them. On Frontière AI, EU sovereign: Llama 3.3 70B — Scaleway (France), fallback OVHcloud (France). Fast access only, served by US companies and labeled as such: Muse Spark 1.1 and Muse Spark 1.2. Llama 3.3 70B is one of the most common models on European price lists: on 2026-10-13, OVHcloud, IONOS and STACKIT all listed it.

gpt-oss (OpenAI)

gpt-oss is OpenAI's open-weight family, under Apache-2.0. Because the weights are public, a European company can run it — unlike OpenAI's closed models, covered below. On Frontière AI, EU sovereign: GPT-OSS 120B — Scaleway (France), fallback OVHcloud (France). STACKIT also listed gpt-oss on 2026-10-13. Nebius serves it too, but its public endpoints carry no region commitment.

Gemma (Google)

Gemma is Google's open-weight family. Gemma 4 is published under Apache-2.0; Gemma 3 under Google's Gemma Terms of Use. On Frontière AI, EU sovereign: Gemma 4 26B — AKI.IO (Germany). Fast access only, served by US companies and labeled as such: Gemma 3 27B. On 2026-10-13, Gemma 4 was also served by T-Systems (Germany) and Infomaniak (Switzerland), and STACKIT listed Gemma models.

Kimi (Moonshot AI)

Moonshot AI publishes Kimi weights under its own licence. Few providers serve Kimi K3 yet. On Frontière AI: no sovereign version of this family today. Fast access only, served by US companies and labeled as such: Kimi K3. In Europe on 2026-10-13: Kimi K3 at Berget AI (Sweden) and TensorX (Ireland, Finland); Kimi K2.6 at Infomaniak (Switzerland).

Apertus (Swiss AI Initiative)

Apertus comes from the Swiss AI Initiative (EPFL, ETH Zurich, CSCS). It is fully open — weights, code and documented training data — under Apache-2.0: a European-built model, not only a European-hosted one. On Frontière AI, EU sovereign: Apertus v1.5 70B — Infomaniak (Switzerland). Infomaniak serves it from its own data centres in Switzerland.

LLMs hosted in France

If your requirement is narrower — not just the EU, but France — the models whose primary route on Frontière AI is a French provider (Scaleway and Mistral AI) are:

• Qwen3 235B — Scaleway (France)
• Qwen3.5 397B — Scaleway (France)
• GLM-5.2 — Scaleway (France)
• DeepSeek V4 Flash 0731 — Scaleway (France)
• Llama 3.3 70B — Scaleway (France), fallback OVHcloud (France)
• Qwen3.6 35B A3B — Scaleway (France)
• Qwen3 Coder 30B — Scaleway (France)
• Pixtral 12B — Scaleway (France)
• GPT-OSS 120B — Scaleway (France), fallback OVHcloud (France)
• Mistral Small 3.2 24B — Scaleway (France), fallback OVHcloud (France)
• Mistral Large 4 — Mistral AI (France)
• Qwen3.8 27B — Scaleway (France), fallback OVHcloud (France)
• Mistral Small 4 119B — Mistral AI (France)

OVHcloud runs its AI Endpoints in Gravelines and Scaleway runs its Generative APIs in Paris only; Mistral's EU endpoint covers EU and EFTA data centres without naming a single country. When a fallback is listed, it is another European sovereign provider, which may be in another EU country. Scaleway does not store prompts, except requests that cause errors or abuse, which it may keep for up to two weeks.

Closed models: EU-hosted, US-controlled

The best-known closed models can also run in Europe, through their US vendors. The three largest each describe it differently on their own documentation, read on 2026-10-19:

• OpenAI offers a Europe region (EEA and Switzerland) on eu.api.openai.com, with in-region storage and in-region inference for the endpoints it lists. It requires Zero Data Retention or Modified Abuse Monitoring (or one of their variants), and it does not cover "system data", which may be processed and stored outside the region (OpenAI docs).
• Anthropic's own API offers two inference geographies, global by default and us; workspace storage is US only. On Amazon Bedrock and Google Cloud, the region follows the endpoint or inference profile you pick (Claude docs).
• Google Vertex AI keeps machine-learning processing within the jurisdiction of the location-specific endpoint you call; its eu endpoint covers EU member states, not the UK or Switzerland (Google Cloud docs).

All three are useful answers to "where does inference run?". None changes who controls the service: the operator remains a US company, within reach of the CLOUD Act wherever its servers are. Whether that is acceptable depends on your data and your contracts; our GDPR vs. CLOUD Act guide walks through the analysis. If it is not, the open-weight families above are the alternative.

How to verify a "hosted in the EU" claim

"Hosted in Europe" on a product page is a starting point. Before you send personal or confidential data, check:

1. The legal notice and the parent company. Who operates the service, where it is incorporated, and who controls it. A European region inside a US company answers where, not who.
2. The region commitment. Look for a commitment in the terms or legal guide, not just a map. Some providers route "dynamically" and say the region can change without notice.
3. The sub-processor list. It shows whether a third party — sometimes outside Europe — touches the request.
4. The retention sentence, word for word. Zero retention by default, zero retention with an exception for errors or abuse, or 30 days unless you ask: three different answers.
5. The training clause. Including exceptions for preview or beta models, like Mistral's § 4.3.
6. Per-request evidence. Can you see, for each call, which provider served it? Without that, a fallback to another region is invisible.

Our sovereignty methodology applies these checks to every provider we route to, and each model page shows the result.

How to call EU-hosted models with one API

Nearly every provider in this article speaks the OpenAI format, so switching is a base URL, a key and a model ID. Through Frontière AI, all the sovereign models above sit behind one OpenAI-compatible endpoint. The X-Data-Sensitivity: sensitive header makes the gateway refuse any non-sovereign route for that request — a key can enforce the same rule permanently:

curl https://getfrontiereai.eu/api/v1/chat/completions \
  -H "Authorization: Bearer $FRONTIERE_API_KEY" \
  -H "Content-Type: application/json" \
  -H "X-Data-Sensitivity: sensitive" \
  -d '{"model": "glm-5.3",
       "messages": [{"role": "user", "content": "Summarise this contract clause..."}]}'

Each response carries X-Routed-Provider (the provider that served it) and X-Sovereign (true or false), so your own logs prove where each request ran. If no sovereign provider is available for the model, the call fails with an error instead of leaving Europe. The documentation covers the Python and JavaScript SDKs, and the EU LLM gateway guide explains routing and failover in detail.

Where Frontière AI fits

Frontière AI is a European LLM gateway published by FLEECE AI SASU, a French company. It does not own GPUs: it routes to the European providers named above and labels every model by jurisdiction. Prompts and completions are not written to our database; the gateway itself runs on Vercel's Paris region — a US company, in transit only — as our privacy policy states. Billing is prepaid from €10, VAT included, with no subscription, for companies and individuals.

Buying directly from one provider is cheaper, and if one provider covers your models, you should. A gateway is worth it when you want several families — say DeepSeek from Germany, Qwen from Paris and Apertus from Switzerland — behind one key, with the jurisdiction named on every call. Browse the full catalogue.

FAQ