A tinted zone holding three blocks, one purple-topped, with one dashed outline left outside it, suggesting models served inside a boundary and one retired model outside.

Why 'EU provider' and 'EU-hosted model' are different claims

Where you can run an open source LLM in Europe depends on the model and the provider together. An open-weight model is one whose trained weights are published, so anyone can download and run it. Search engines call these open source LLMs, and we mean the same models here. Moonshot, for example, releases the full Kimi K3 weights "for research, deployment, and further innovation". So the model name alone tells you nothing about where a given provider runs it.

Which open models can you call in Europe today, and how do you prove where each one runs? The question usually arrives in a security questionnaire, with one field per model you use. A provider's company address does not fill that field.

Lyceum is a sovereign European AI inference provider for open-source models. Our serverless models run in Paris and Finland, and every model page states its processing region.

We will follow one hypothetical example through the article. A support team summarises 20,000 tickets a month with one open model. Its data protection officer (DPO), the person who answers for personal data in the company, wants one line per model in the record.

The model-to-region table

The table lists the open models on our models page, with the processing region each model's page shows. We read every page on 23 October 2026.

ModelAPI model IDProcessing region
DeepSeek V4.1 Flashdeepseek/deepseek-v4.1-flashEU-hosted
DeepSeek V4 Flashdeepseek/deepseek-v4-flash-0731EU-hosted
DeepSeek V4 Pro 0813deepseek/deepseek-v4-pro-0813EU-hosted
GLM-5.3z-ai/glm-5.3EU-hosted
GLM-5.3 Flashz-ai/glm-5.3-flashEU-hosted
GLM-5.2z-ai/glm-5.2EU-hosted
Kimi K3moonshotai/kimi-k3EU-hosted
Kimi K2.7 Codemoonshotai/kimi-k2.7-codeEU-hosted
MiniMax M3minimax/minimax-m3EU-hosted
Qwen3.8 2.4T A95Bqwen/qwen3.8-2.4t-a95bEU-hosted
Qwen3.8 27Bqwen/qwen3.8-27bEU-hosted
Qwen3.8 Flash Nextqwen/qwen3.8-flash-nextEU-hosted
Qwen3 Embedding 8Bqwen/qwen3-embedding-8bEU-hosted

Model names follow the publisher. In Qwen3.8 2.4T A95B, 2.4T is the total parameter count and A95B means 95B parameters are active per token.

The instant variants of GLM and Qwen are the same models with the reasoning phase switched off. Their pages show EU-hosted too. So do our routing aliases, such as lyceum/router. An alias chooses a listed model for each request and bills it as that model.

For the example team, the row that matters is MiniMax M3. Its page lists $0.40 per 1M input tokens and $2.00 per 1M output tokens. Say each ticket sends 3,000 input tokens and gets a 400-token summary back:

monthly cost = tickets x (input tokens x input price + output tokens x output price)
             = 20,000 x (3,000 x $0.40 + 400 x $2.00) / 1,000,000
             = $40.00

That is a calculated example at list prices of 23 October 2026. It assumes no cached input. Cached input is a prompt start that repeats a recent request and is still in GPU memory. Those tokens cost $0.10 per 1M instead of $0.40, so a team that starts every prompt with the same instructions pays less. The example does not show other models or real ticket lengths.

What global routing means for a residency answer

Some providers label a model's region "Global" instead of naming a place. For a residency answer, the question is whether such a request can be processed outside the EEA, the EU plus Iceland, Liechtenstein and Norway.

The European Data Protection Board (EDPB), the body of EU data protection authorities, names three criteria that must all apply for a transfer. The sender is subject to the GDPR. It makes personal data available to another organisation. That organisation is in a country outside the EEA.

A transfer is allowed, but only under Chapter V of the GDPR, its rules for sending personal data outside the EEA. The first route is an adequacy decision: the European Commission finds that a country protects data about as well as the EEA. The second is appropriate safeguards, such as standard contractual clauses (SCCs), the Commission's model contract for transfers. A few narrow exceptions exist.

SCCs require a transfer impact assessment. You, as the sender, document the transfer, the laws of the destination country and the extra safeguards. A region field that reads "Global" can therefore turn a one-line answer into that paperwork. Our guide to AI data residency covers storage and processing in depth.

Two routes from your app. On the top route the model page shows EU-hosted, our serverless models run in Paris and Finland, and the record needs one line with the model ID, region and read date. On the bottom route the region reads Global, the location is chosen per request, and if it can be outside the EEA the record needs a transfer check under Chapter V of the GDPR, such as standard contractual clauses with a transfer impact assessment. The transfer check is highlighted.
An EU-hosted model needs one record line, while a Global region that can leave the EEA needs a transfer check.

How to verify a provider's claim yourself

You can check a residency claim in four steps, with any provider and with us.

  1. List the model IDs your code sends. Our `GET /models` endpoint is the authoritative list of IDs the API accepts right now.
  2. Open each model's own page and read its Processing region field. The page itself says: "Check the model's processing region before sending data."
  3. Read what happens to the prompt. Our trust page says we don't keep your prompts or outputs and never use them for training. A request leaves only a usage record: organisation and user, model, token counts and time.
  4. Ask for the documents. Our DPA, the data processing agreement that sets how we handle your personal data, is Annex 1 of our terms. Subprocessors are the companies that run part of our service for us. Their list is available on request by email.

Step 1 takes one request with your API key:

curl https://api.lyceum.technology/openai/v1/models \
  -H "Authorization: Bearer $LYCEUM_API_KEY"

For the example team, step 1 returns the ID it uses, minimax/minimax-m3. Step 2 shows EU-hosted on 23 October 2026.

What to put in your record of processing activities

Under our DPA, you are the controller, the party that decides why personal data is processed. We are the processor, which processes it on your behalf. Article 30 of the GDPR makes each controller keep a record of processing activities. The processor's own record must name any transfer to a third country. So your entry for a model can point to the region on its page, with no transfer to declare when it reads EU-hosted.

For one model, the entry can look like this, filled in for the example team:

FieldExample entry
PurposeSummarising customer support tickets
Model and IDMiniMax M3, minimax/minimax-m3
Processing regionEU-hosted, per the model page, read 23 October 2026
GPU locationsParis and Finland, for serverless models
RetentionPrompts and outputs not kept; usage record only
ContractDPA, Annex 1 of our terms
Volume and cost20,000 tickets a month, $40.00 at list prices of 23 October 2026
Four arrows link where each fact comes from to the record line it fills, for a hypothetical support team on MiniMax M3. GET /models fills model and ID. The model page's Processing region field fills region and read date, which is highlighted. The trust page fills retention. The DPA, Annex 1 of our terms, fills the contract line.
For the hypothetical support team on MiniMax M3, each record line comes from one of our pages, read on 23 October 2026.

Our ISO 27001 audit is in progress, so the record should not name a certificate of ours. The data centre operators hold ISO certifications for their own facilities.

A routing alias needs one more line. The router chooses per request, and its page names the model it sends traffic to at present. On 23 October 2026 that was MiniMax M3. For a regulated workload we recommend a fixed model ID, so the record names one model.

Retired models to move off if residency is absolute

A retired model is one we no longer serve. Retirement says nothing about where the model used to run. For a residency file, the point is that a retired ID has no current model page to read a region from. Our models page lists each one with the model to use instead. If your code or your record still names one, update both.

Retired modelFormer model IDUse instead
MiniMax M2.5minimax/minimax-m2.5MiniMax M3
Kimi K2.6moonshotai/kimi-k2.6Kimi K3
GLM-5.1z-ai/glm-5.1GLM-5.3
Qwen3.5 397B A17Bqwen/qwen3.5-397b-a17bGLM-5.3 Flash
Qwen3.5 9Bqwen/qwen3.5-9bQwen3.8 Flash Next
Nemotron 3 Ultra 550Bnvidia/nemotron-3-ultra-550b-a55bDeepSeek V4 Pro 0813
Nemotron 3 Super 120Bnvidia/nemotron-3-super-120b-a12bGLM-5.3 Flash
Gemma 3 27Bgoogle/gemma-3-27b-itGLM-5.3 Flash

The example team used minimax/minimax-m2.5 until it was retired. It changes one model string in its code. Then it reads the MiniMax M3 page and updates the record.

Before your next review, check your model against the table above. Then copy its region and the read date from its page into your record.

Sources

  1. huggingface.co
  2. edpb.europa.eu
  3. gdpr-info.eu