Which AI models run in Europe, and who runs them?
DeepSeek, Qwen, Mistral and five other open models that twelve providers sell by the token: where the model actually runs, what a million tokens cost, what the provider may do with your prompts, and whose law applies to the provider. Fetched from the providers’ own catalogues every month and worked out by us, not by a provider.
All eight open models run in their newest release at a European provider in Europe. At the hyperscalers, four of them run in a European region.
Where a European provider and a hyperscaler run the same release in Europe, the European one is cheaper for one in three: gpt-oss 120B at OVHcloud. For Gemma 31B and Nemotron 3 Super AWS is cheaper. The price weighs input and output 3:1, as Artificial Analysis does, and compares the release each provider’s cell shows.
October 2026 edition · prices fetched 29 October 2026
What does a million tokens cost, input and output?
List price as an index: 100 is the cheapest provider in Europe of the same release. The upper bar is input, the lower output. Excluding VAT.List price in euros per million tokens, excluding VAT. The upper figure is input, the lower output. The bars are the same: 100 is the cheapest provider in Europe of the same release.
We have room for eight models. In are the eight most demanded open models that at least one of the providers sells by the token, in order of demand. That is our reading of three numbers: tokens on OpenRouter and downloads on Hugging Face over the last month, and how many of the providers run the model in Europe. It says nothing about which model is best.
| Provider | DeepSeek1V4.1 Flash · V4 Pro | Gemma24 31B · 4 26B | GLM35.3 Flash · 5.3 | Qwen43.8 27B · 3.8 Flash Next | gpt-oss5120B · 20B | Nemotron63 Super · 3 Ultra | Mistral7Small 4 · Medium 3.5 · Large 3 | Kimi8K3 | |
|---|---|---|---|---|---|---|---|---|---|
| AzureUS | global V4.1 Flash · V4 Flash in EU zone | – | global 5.3 Flash | – | unclear | global 3 Ultra · US zone | other release0.481.45 Large 3 · EU zone · preview | global K3 | |
| GoogleUS | global V4.1 Flash | global 26B | – | – | global 120B | – | older release0.090.27 Small 3.1 · Netherlands | – | |
| AWSUS | older release0.660.331.991.00 V3.2 · Stockholm | 1000.150.08 1000.430.22 31B · Frankfurt · no cache | global 5.3 · GLM 5 in Europe | older release0.130.070.540.27 Qwen3 32B · Stockholm | 1680.130.07 1350.540.27 120B · Stockholm · no cache | 1000.160.08 1000.700.35 3 Super · Stockholm · no cache | global Large 3 · US zone · Devstral 2 in Europe | global K3 · K2.5 in Europe | |
| ScalewayFR | 1000.600.360.30 1001.800.90 V4.1 Flash · preview | other release0.250.130.500.25 26B · preview | other release1.800.905.502.75 5.3 · preview | 1501250.600.360.30 5082543.301.65 3.8 27B · preview | 1880.150.08 1500.600.30 120B · no cache | – | other release1.500.757.503.75 Medium 3.5 · preview | – | |
| OVHcloudFR | – | – | – | 1001390.400.20 4152082.701.35 3.8 27B · FP8 · no cache | 1000.080.04 1000.400.20 120B · FP4 · no cache | – | older release0.090.050.280.14 Small 3.2 · until 4 Nov | – | |
| IONOSDE | – | – | – | 1001392000.40 4152.70 3.8 27B · INT4 · no cache · no batch | 1883750.15 1633250.65 120B · MXFP4 · no cache · no batch | – | older release0.100.30 Small 3.2 | – | |
| STACKITDE | – | 2985960.45 1513010.65 31B · 16-bit · no cache · no batch | – | 1131572250.45 1000.65 3.8 27B · 16-bit · no cache · no batch | 56311250.45 1633250.65 120B · MXFP4 · no cache · no batch | – | – | – | |
| T-SystemsDE · from €1,000/mo | – | 3972297950.600.35 2785571.20 31B · FP8 · no batch | 1000.200.12 1200.60 5.3 Flash · no batch | 1251002500.500.29 3852.50 3.8 27B · FP8 · no batch | 2501445000.200.12 1633250.65 120B · no batch | 3712137420.600.35 1713421.20 3 Super · FP8 · no batch | 2771590.600.35 1481.20 Small 4 · no batch | – | |
| BergetSE | – | 1663310.25 1162320.50 31B · BF16 · no cache · no batch | 1252170.25 1000.50 5.3 Flash · FP8 · no cache · no batch | – | retired 11 Sept | – | older release0.300.30 Small 3.2 | 1003.001.95 10015.00 K3 → K3.1 Pro 6 Nov · until 6 Nov · MXFP4 · no batch | |
| RegoloIT | – | 2655300.40 4879742.10 31B · FP16 · no cache · no batch | older release2.005.20 5.2 | 1251742500.50 3232.10 3.8 27B · NVFP4 · no cache · no batch | 125025001.00 105021004.20 120B · MXFP4 · no cache · no batch | – | 2310.50 2582.10 Small 4 · no cache · no batch | – | |
| InfomaniakCH | – | 1442870.22 1012010.43 31B · no cache · no batch | – | 1081512170.43 5003.25 3.8 27B · FP8 · no cache · no batch | – | – | 1000.22 1000.81 Small 4 · no cache · no batch | older release0.653.25 K2.6 | |
| MistralFR · the maker’s own; EU endpoint +10% | EU +10%, unconfirmed0.300.150.300.151.191.19 V4.1 Flash · preview | – | EU +10%, unconfirmed0.150.090.150.090.490.49 5.3 Flash | – | – | – | EU +10%, unconfirmed0.150.080.150.080.590.59 Small 4 | – | |
DeepSeek1 V4.1 Flash · V4 Pro
DeepSeek, China
- Azure global V4.1 Flash · V4 Flash in EU zone
- Google global V4.1 Flash
- AWS older release0.660.331.991.00 V3.2 · Stockholm
- Mistral EU +10%, unconfirmed0.300.150.300.151.191.19 V4.1 Flash · preview
Not offered at OVHcloud, IONOS, STACKIT, T-Systems, Berget, Regolo, Infomaniak.
Gemma2 4 31B · 4 26B
Google, US
- Google global 26B
- Scaleway other release0.250.130.500.25 26B · preview
Not offered at Azure, OVHcloud, IONOS, Mistral.
GLM3 5.3 Flash · 5.3
Z.ai, China
- Azure global 5.3 Flash
- AWS global 5.3 · GLM 5 in Europe
- Scaleway other release1.800.905.502.75 5.3 · preview
- Regolo older release2.005.20 5.2
- Mistral EU +10%, unconfirmed0.150.090.150.090.490.49 5.3 Flash
Not offered at Google, OVHcloud, IONOS, STACKIT, Infomaniak.
Qwen4 3.8 27B · 3.8 Flash Next
Alibaba, China
- AWS older release0.130.070.540.27 Qwen3 32B · Stockholm
Not offered at Azure, Google, Berget, Mistral.
gpt-oss5 120B · 20B
OpenAI, US
- Azure unclear
- Google global 120B
- Berget retired 11 Sept
Not offered at Infomaniak, Mistral.
Nemotron6 3 Super · 3 Ultra
NVIDIA, US
- Azure global 3 Ultra · US zone
- none
Not offered at Google, Scaleway, OVHcloud, IONOS, STACKIT, Berget, Regolo, Infomaniak, Mistral.
Mistral7 Small 4 · Medium 3.5 · Large 3
Mistral AI, France
- Azure other release0.481.45 Large 3 · EU zone · preview
- Google older release0.090.27 Small 3.1 · Netherlands
- AWS global Large 3 · US zone · Devstral 2 in Europe
- Scaleway other release1.500.757.503.75 Medium 3.5 · preview
- OVHcloud older release0.090.050.280.14 Small 3.2 · until 4 Nov
- IONOS older release0.100.30 Small 3.2
- Berget older release0.300.30 Small 3.2
- Mistral EU +10%, unconfirmed0.150.080.150.080.590.59 Small 4
Not offered at STACKIT.
Kimi8 K3
Moonshot AI, China
- Azure global K3
- AWS global K3 · K2.5 in Europe
- none
- Infomaniak older release0.653.25 K2.6
Not offered at Google, Scaleway, OVHcloud, IONOS, STACKIT, T-Systems, Regolo, Mistral.
Hyperscalers: Microsoft Azure, Google Cloud and AWS, in their European regions.European clouds: European cloud providers that also sell AI models by the token.AI specialists: providers that mainly sell AI inference, and Mistral’s own platform.
- DeepSeek: DeepSeek, China · MIT
- Gemma: Google, US · Apache 2.0
- GLM: Z.ai, China · Flash: MIT · 5.3: a host with over $10 billion in revenue over 12 months must pass Z.ai’s security review
- Qwen: Alibaba, China · 27B: Apache 2.0 · Flash Next: anyone selling it as a service needs a separate licence from Qwen
- gpt-oss: OpenAI, US · Apache 2.0
- Nemotron: NVIDIA, US · 3 Super: NVIDIA Nemotron Open Model License · 3 Ultra: OpenMDW 1.1 · no revenue limit
- Mistral: Mistral AI, France · Small 4, Large 3: Apache 2.0 · Medium 3.5: no rights for a company with over $20 million in monthly revenue
- Kimi: Moonshot AI, China · a host with over $20 million in revenue over 12 months needs a separate agreement with Moonshot, unless it is a certified partner
The licences were read in the models’ official repositories on Hugging Face on 30 October 2026.
T-Systems: only with a plan from €1,000 a month. Europe: the EU, the EEA, Switzerland and the UK. Mistral: the EU endpoint guarantees inference in the EU and EFTA for 10% more, but which models it serves is not public; hence “unconfirmed”, with no bar and outside the count, until Mistral confirms the list.
- The index compares the same release at every provider, also where they run it at different precisions; the precision is in the cell, and lower precision can mean slightly lower quality.
- The price per token is only the price: not speed, usage limits, context length or SLA.
- 100 is the cheapest in Europe of the same release. It may be a preview or a model on its way out.
- Each column has a main release; other releases are shown as text. Mistral’s EU price is our calculation: list price times 1.1.
- Prices in the provider’s own currency, converted at the ECB reference rate. The region is the one nearest Denmark that sells the model.
- The numbers are refreshed on the first Monday of every month. If a provider changes a price in between, it is in the next release.
Our interest: Nomadics helps companies move to European cloud and build AI on it, runs on Scaleway itself and has applied for partnerships with Scaleway and OVHcloud. That is why we include the hyperscalers in their European regions and let the price stand, also where an American provider is cheapest. No provider pays to be included, and none saw the table before it was published. If we receive a referral fee from a provider a client chooses, the offer says so.
What is new in this release
- First release. Eight open models at twelve providers; the catalogues were fetched on 29 October 2026, and the terms read on 29 and 30 October.
- DeepSeek V4.1 Flash is the most used open model on OpenRouter in October; in Europe it runs at Scaleway, and at Azure and Google only on a global endpoint.
- Berget took Qwen 3.8 27B out of its catalogue on 29 October; Qwen runs in Europe at seven other providers.
- Mistral’s EU endpoint does not count until Mistral has confirmed which models it serves; Mistral was asked on 29 October.
- Next release: Monday 2 November 2026, then the first Monday of every month.
The numbers as data
The whole table for October 2026 – every source, every model, every price in the provider’s currency and the ECB rate – is available as JSON: this month (2026-10.json) · the newest release (ai-modeller.json)
The numbers may be used freely with attribution (CC BY 4.0). Cite as: Nomadics, “AI models in Europe, October 2026”, nomadics.dk/en/method/ai-models/.
The providers’ own catalogues and price lists the numbers were calculated from are kept unchanged for every month, so a price can be checked even after the provider has replaced it.
All releases
Every month gets a page of its own, and the newest is this one.
- October 2026 · prices fetched 29 October 2026 · this page
What the provider says about your prompts
One row per provider: the provider’s own words on what happens to your prompts and outputs, quoted verbatim from the document with its date, where the inference runs, and which company you contract with. The exceptions are in the quote, not in a summary.
| Provider | On your prompts | Where it runs | Whose law |
|---|---|---|---|
| Azure Hyperscalers · US | “Your prompts (inputs) and completions (outputs), your embeddings, and your training data: are NOT available to other customers. are NOT available to OpenAI or other providers of Models sold by Azure. are NOT used to train any generative AI foundation models without your permission or instruction.” Data, privacy, and security for Foundry Models sold by Azure, 5 June 2026 · read 29 October 2026 The same page: models in preview “may employ different privacy practices, including with respect to abuse monitoring”. Most open models at Azure are in preview. | EU data zone (Azure's EU Data Boundary, which can include Norway and Switzerland) for the models the table shows in Europe; the rest global only | Microsoft Ireland Operations Ltd, Dublin; parent Microsoft Corporation, US |
| Google Hyperscalers · US | “Google won't use your data to train or fine-tune any AI/ML models without your prior permission or instruction. This applies to all managed models on Gemini Enterprise Agent Platform, including GA and pre-GA models.” Gemini Enterprise Agent Platform and zero data retention, 28 October 2026 · read 29 October 2026 Google Cloud Platform Terms, 12 October 2026, section 4.3: where Google's tools detect potential abuse, “Google may log Customer prompts solely for the purpose of reviewing and determining whether a violation has occurred”. | Global endpoint for the open models: “doesn't support data residency requirements”; only Mistral's partner models in europe-west4 (the Netherlands) | Google Cloud EMEA Ltd, Dublin; parent Google LLC, US |
| AWS Hyperscalers · US | “With Amazon Bedrock, your content is not used to improve the base models and is not shared with any model providers.” Amazon Bedrock FAQs, undated · read 29 October 2026 The user guide, Data retention: in mode none “No request or response data is written to durable storage by AWS or shared with the model provider”; in mode default “AWS may retain the data for safety and abuse-prevention purposes”. | In the region the model is called in (Stockholm, Frankfurt, Ireland, Milan, London); Kimi K3 and GLM 5.3 on a global endpoint only | Amazon Web Services EMEA SARL, Luxembourg; parent Amazon.com, Inc., US |
| Scaleway European clouds · FR | “Scaleway does not retain any requests or generated Content after processing.” AI Services Specific Conditions, section 4.4, version of 7 April 2026, 7 April 2026 · read 29 October 2026 The documentation's privacy page (reviewed 3 October 2025): “By default we apply a Zero Data Retention Policy”; on misuse or errors that harm the service, “Data is stored for up to two weeks, only to reproduce, investigate, and fix the underlying issue”; “Your data is not used for training, retraining, or improving the base models.” (the document) | Paris, France (a data centre run by OPCORE) | Scaleway SAS, Paris (iliad group) |
| OVHcloud European clouds · FR | “OVHcloud has no knowledge of Inputs and Outputs, and OVHcloud does not reuse them in any way. OVHcloud does not carry out any backups of Inputs or Outputs. This does not apply for the AI Endpoint Batch API Service.” Specific Conditions for Public Cloud Services, Appendix 10: AI Endpoints, 26 August 2026 · read 29 October 2026 | “The European Union” (the product page); the provider does not name the place | OVH SAS, Roubaix, France (OVH Groupe SA) |
| IONOS European clouds · DE | “The content of your prompts and of the outputs generated for you is never logged and is never accessed by IONOS Cloud personnel. It is not written to any log or to persistent storage; it is not used to train, fine-tune or otherwise improve any model; and it is not disclosed to any third party, including the developers of the models we operate.” AI Model Hub, Data handling (documentation), undated · read 29 October 2026 The same page: “No third-party sub-processors are engaged for AI Model Hub.” The terms for KI-Dienste (section 2.3) allow improving the service from inputs and outputs only “bei der Nutzung von SaaS”; the Model Hub is API access to models, PaaS under section 1.1 of the terms. | “IONOS Cloud data centers in Germany” | IONOS Cloud GmbH, Montabaur, Germany (IONOS Group SE) |
| STACKIT European clouds · DE | “When using the AI Model Serving Service, STACKIT does not collect or evaluate any customer data other than billing-relevant data. A backup of Customer requests does not take place.” Service Certificate STACKIT AI Model Serving, version 1.4, valid from 6 November 2026, 6 November 2026 · read 29 October 2026 The annex: “The e-mail and subject ID are stored temporarily (30 days) in the logs of the service.” | Region EU01, “Germany South” | Schwarz Digits Cloud GmbH & Co. KG, Bad Friedrichshall, Germany (Schwarz Group) |
| T-Systems European clouds · DE | “By default, your requests and the responses received are not permanently stored in the LLM Serving Service, nor are they used to train the language models. To make the service respond faster, there is a so-called caching function: already processed requests are temporarily stored for a limited time so that similar or recurring requests can be answered more quickly. This function is enabled by default but can be switched on or off by customers at any time.” Service Specifications and Additional Terms and Conditions – LLM Serving Service, version 1.24, section 2.3.1 (the English text is a convenience translation; the German prevails), 30 September 2026 · read 29 October 2026 | Germany for the models Telekom runs itself (the ones in the table); the contract promises “within the EU/EEA and/or in countries with an adequacy decision” | T-Systems International GmbH, Bonn (Deutsche Telekom AG) |
| Berget AI specialists · SE | “Berget AI applies a true Zero Data Retention policy on its serverless inference services, including Berget Code. Berget AI does not train any models on the input or output from serverless inference endpoints, nor does it share the data with any model providers for model training.” Data Processing Agreement, 22 October 2026 · read 29 October 2026 | Sweden, on Berget's own GPUs | Berget AI AB, Stockholm |
| Regolo AI specialists · IT | “Your data is never stored or reused, ensuring compliance with GDPR and beyond.” regolo.ai, the front page, undated · read 30 October 2026 The privacy policy, retention of inputs and outputs: “Never, excepted the use of storage services activated by the user.” (the document) | “EU Data Center”; the provider does not name the country | Seeweb S.r.l., Frosinone, Italy (DHH group) |
| Infomaniak AI specialists · CH | “Your queries are neither recorded nor used to train models or improve our services.” AI Services, the product page, undated · read 30 October 2026 | Switzerland, in Infomaniak's own data centres | Infomaniak Network SA, Geneva, Switzerland |
| Mistral AI specialists · FR | “Except for specific APIs, we keep your Input and Output for the period necessary to generate the Output and then for thirty (30) rolling days to monitor abuse (unless zero data retention is activated).” Privacy Policy, 3 September 2026 · read 29 October 2026 Commercial Terms of Service, 25 September 2026, section 4.3: “By using Labs or Preview Models, you acknowledge that (i) Mistral AI may use Customer Data and Outputs generated from Labs or Preview Models to train its artificial intelligence models”. The sub-processor list (Trust Center) names Microsoft (Sweden, Norway) and CoreWeave (EEA) as infrastructure for the API. (the document) | Global endpoint with no committed place; the EU endpoint (+10 %) promises “Multiple data centers in EU and EFTA countries”, partly on American companies' infrastructure | Mistral AI, Paris |
The company that sells the service and its parent where it has one; not the data centre’s country.
Every quote can be checked: the link leads to the document. When a provider moves a sentence, we read the document again and correct the row in the next release.
The rules
- A model is in when the provider sells it by the token through an API and runs it itself. GPUs you rent and run yourself are not in.
- “In Europe” means the provider promises in writing that inference happens in the EU, the EEA, Switzerland or the United Kingdom: a region or an EU data zone. A global endpoint is shown as “global” and does not count. Where a provider’s terms say “EU”, we quote them as they stand.
- List prices per million tokens, input and output, excluding VAT, in the provider’s currency and converted to euros and kroner at the ECB reference rate on the day the prices were fetched. Mistral’s EU endpoint costs list price × 1.1, and that is the price shown.
- For each model, the newest release the provider sells. Older releases are shown as text and do not count in the headline.
- The terms are quoted verbatim with the document’s name and date. We do not summarise them.
- “Whose law” is the company that sells the service and its parent, not the data centre’s country.
- We say nothing about which model is best. The benchmark sites do that.
- T-Systems sells by the token only with a plan from €1,000 a month, and the provider says so.
- We have room for eight models: the most demanded open models that at least one of the providers sells by the token, in order of demand. Demand is the average of three shares: the model’s share of tokens on OpenRouter and of downloads on Hugging Face in the last month, and its share of the providers that run it in Europe. A new model takes the eighth place once it has scored above it two months in a row.
- OpenRouter publishes only its top 20, so a model below it counts as zero there; OpenRouter is one router, skewed to developers and coding agents; a model maker’s own platform publishes no numbers; downloads count repeat users and pipelines. The score says which way demand leans, not market share.
Our interest
Nomadics helps companies move to European cloud and build AI on it, runs on Scaleway itself and has applied for partnership with Scaleway and OVHcloud. So we quote the providers’ terms verbatim and with their date, count the hyperscalers in their European regions and let the price stand, even where an American provider is cheapest. No provider pays to be included, and none saw the table before it was published. If we receive a referral commission from a provider a client chooses, the proposal says so.
Why it matters
Among EU enterprises that have considered AI, lack of clarity about the legal consequences is the second most cited reason for holding back (52.52 %), and concerns about data protection the third (48.83 %). Denmark has the highest share of enterprises using AI: 42.03 %.
Eurostat, Use of artificial intelligence in enterprises (data extracted December 2025)
Every month
We fetch the catalogues again on the first Monday of every month with the same rules. A program reads the providers’ own catalogues and price lists and recalculates everything; the date is under the figure. If the numbers are more than a month old, the page says so.
We price the cloud itself the same way, every quarter
Your own models
Which model and which provider fit your cases depends on data, language and what has to be approved. The assessment works that out on your own numbers.