ModelsOpen weights

Running DeepSeek models on European infrastructure

DeepSeek's open weights run on European infrastructure. See which EU providers serve which DeepSeek models and what each states about hosting.

, AI Researcher5 min readUpdated 1 October 2026
In this article7
  1. Where a model was trained is not where it runs
  2. What the DeepSeek weights permit
  3. Which European providers serve which DeepSeek models
  4. Checking the hosting region per model
  5. Choosing a European route to DeepSeek
  6. Frequently asked questions
  7. Sources
Diagram separating three facts about an AI model: where it was trained, who publishes its weights, and the hosting region of the endpoint serving it
Diagram separating three facts about an AI model: where it was trained, who publishes its weights, and the hosting region of the endpoint serving it

Where a model was trained is not where it runs

DeepSeek is a Chinese AI developer based in Hangzhou, which leads many readers to assume that using its models means sending prompts to China.[1] That conflates three separate facts: where a model was trained, who published its weights, and where an endpoint runs. Only the third determines where a prompt travels.

  • Training location: where the compute that produced the model sat. It is fixed once training ends and says nothing about later use.
  • Weight publication: the developer releases the trained parameter files under a licence. The DeepSeek V4.1 Flash model card states that the repository and the model weights are licensed under the MIT License, which permits commercial use and self-hosting.[2]
  • Endpoint location: the data centre that actually processes a prompt. This is decided by whoever serves the weights, not by who trained them.

What the DeepSeek weights permit

Open weights means downloadable parameter files that anyone can load and serve on their own hardware. The consequence is practical: a DeepSeek model served from a European endpoint processes European prompts there, and the vendor's own API is one option among several, not the only route. What matters for data location is the serving provider's stated hosting region for that specific model.

Which European providers serve which DeepSeek models

A reader who wants DeepSeek without running it themselves can call it from Lyceum's per-token API: DeepSeek V4.1 Flash at $0.50 per 1M input and $1.50 per 1M output tokens, USD excl. VAT (read 2026-09-22), and DeepSeek V4 Flash. Other European providers list DeepSeek models as well, but their pages differ in what they state about models and regions.

ProviderDeepSeek models servedHosting region stated per model
LyceumDeepSeek V4.1 Flash; DeepSeek V4 Flash–
IONOSNone listed–
NebiusDeepSeek V4.1 Flash; DeepSeek-V4-Flash-0731; DeepSeek-V4-Pro; DeepSeek-V4-Pro-0813V4.1 Flash: us-north1; V4-Flash-0731: us-central1; V4-Pro: uk-south1; V4-Pro-0813: us-north1
Open Telekom CloudNone listed–
OVHcloudNo per-token model API on the prices page–
Scalewaydeepseek-v4-flash-0731Paris
STACKITNone named–
TensorXdeepseek-v4.1-flash; deepseek-v4-pro-0813; deepseek-v4-flash-0731–

A dash means the provider's page lists no value. Nebius's four DeepSeek models sit in us-north1, us-central1 and uk-south1, none of them EU regions, while Scaleway's page names Paris for deepseek-v4-flash-0731.[8] How European providers disclose regions, prices and data processing terms more broadly is compared in a separate review of serverless inference providers.

Checking the hosting region per model

Three checks separate a defensible choice from a marketing claim:

  1. Verify the hosting region on the provider's own model page, for the exact model ID, and record the retrieval date. Catalogue pages and sales material are not a substitute.
  2. Treat a missing region statement as unanswered, not as EU. Absence of a US region is not presence of an EU one.
  3. Compare per-token serverless prices only within the same product mode. Per-token inference and per-second GPU billing price different products and never belong in one comparison.

Stated European regions are the exception in the table, and a region stated for one model says nothing about the next. What EU residency can and cannot promise, including where US jurisdiction reaches, is set out in the guide to data sovereignty.

Verifying residency claims per model, rather than trusting a provider-level statement, is the pattern applied in the guide to data residency verification for enterprise AI.

Choosing a European route to DeepSeek

Which route fits depends on what a team must be able to show about where its prompts are processed.

  1. A stated European region for one named model: Scaleway's page names Paris for deepseek-v4-flash-0731[8].
  2. A per-token API from another provider in the table: treat a "–" as an open question and settle it with that provider before production data is sent.
  3. An endpoint under the team's own control: serve the MIT-licensed weights[2] on rented GPU capacity in an EU region, so the region is a fact the team sets itself.

Frequently asked questions

Can I use DeepSeek without sending data to China?

Yes. DeepSeek publishes its model weights openly, so they can be served by any provider on infrastructure outside China. What determines where your prompts go is the hosting region of the endpoint you call, not where the model was trained or where the vendor is based. Check the provider's own model page for a per-model region statement before committing.

Which European providers host DeepSeek?

As of 2026-09-22, Lyceum (DeepSeek V4.1 Flash and V4 Flash), Scaleway, TensorX and Nebius list DeepSeek models; IONOS, Open Telekom Cloud, OVHcloud and STACKIT list none. Of those that do, only Scaleway states a European region for a DeepSeek model: Paris for deepseek-v4-flash-0731. Nebius serves its four DeepSeek models from us-north1, us-central1 and uk-south1, which are US and UK regions.

Are DeepSeek's weights open?

Yes. DeepSeek publishes downloadable model weights, and the DeepSeek V4.1 Flash model card states that the repository and the model weights are licensed under the MIT License. That licence permits commercial use, copying and self-hosting without a separate agreement with DeepSeek.

Where does a DeepSeek endpoint run?

Wherever the serving provider runs it. Hosting region is a per-model fact: Scaleway states Paris for deepseek-v4-flash-0731, while Nebius states us-north1 for DeepSeek V4.1 Flash. A provider's platform-level claim about EU infrastructure does not answer where one specific model runs; where a page states no region, the answer is simply not published.

Sources

  1. en.wikipedia.org
  2. huggingface.co
  3. lyceum.technology
  4. cloud.ionos.com
  5. tokenfactory.nebius.com
  6. docs.llmhub.t-systems.net
  7. ovhcloud.com
  8. scaleway.com
  9. stackit.com
  10. tensorx.ai

Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score

Related reading

All Models →