Best Open Source AI

Open-source AI in Europe: the models, the rules and where to run them

Europe now has its own open-weight models, its own rules for them and its own reasons to keep AI close to home. A country-by-country guide to the models, the law and the path to an AI a European company controls.

By The editors · Updated · 6 min read

A small team reviews printed maps of data-centre locations at a table in a European city office, with rooftops visible through the window.
For European companies, where a model runs is part of the decision.
Contents
  1. Why the question is different in Europe
  2. The European models
  3. What the law requires
  4. Country by country
  5. The path for a European company
  6. Questions buyers ask
  7. Sources

Why the question is different in Europe

A European company choosing an AI model is choosing more than capability. It is choosing which law applies to the data that passes through the model, who can be compelled to hand that data over, and which obligations follow from training a model of its own. Those questions have sharpened. The US CLOUD Act requires US service providers to disclose data in their "possession, custody, or control", regardless of whether it is stored inside or outside the United States, although providers can move to quash or modify a request⁠1. Transfers of personal data to the US rest on the EU-US Data Privacy Framework, which the EU General Court upheld in September 2025⁠2; the challenger has appealed to the Court of Justice⁠3.

Governments are already acting on this. France's interministerial platform, Albert API, describes itself as the State's sovereign AI infrastructure, built on open-source models and compatible with SecNumCloud hosting⁠4. Spain's digital ministry runs a community for open-source AI⁠5, and Poland's digitisation ministry released eleven Polish-language PLLuM models on open licences in May 2026⁠6.

Open-weight models make an answer possible that rented AI does not: the weights can run on European infrastructure, under European law, with no request leaving the company's control, and they can be trained on the company's own data without that data being sent to a model provider.

The European models

ModelDeveloperLicenceWhat stands out
Mistral Small 4, Ministral 3Mistral AI, FranceApache 2.0⁠7Small 4 has 119B parameters with 6.5B active and a 256K context⁠8
Mistral Large 4Mistral AI, FranceMistral Ecosystem Open-Weight License⁠9About 1T parameters, 52B active; a separate licence is needed to run it as a service⁠109
EuroLLM-22BEU-funded consortium led from LisbonApache 2.0⁠1135 languages, including every official EU language⁠1112
Kolibri-1Aleph Alpha, GermanyApache 2.0 for the weights⁠1378B parameters, 3.46B active, German and English, released 3 October 2026⁠13
Prelude 9BOpenEuroLLMApache 2.0⁠14Experimental and awaiting alignment⁠14
Teuken-7BOpenGPT-X (Fraunhofer and partners)v0.6: CC-BY-NC-4.0, non-commercial⁠1524 EU languages; a research version is published separately⁠1516
PLLuMPolish consortiumOpen licences⁠6Eleven models from 4B to 70B parameters, released 21 May 2026⁠6
GPT-NLTNO, NFI and SURFNot yet publishedVersion 1 due in autumn 2026, initially for launching customers⁠17

OpenEuroLLM, the largest public effort, runs from February 2025 to January 2028 with a budget of €34 million, €20.5 million of it from the EU, and brings together about twenty partners coordinated by Charles University⁠1819. Its first year produced small reference models and Prelude 9B⁠2014. For production work today, Mistral's Apache models, EuroLLM-22B and Kolibri-1 are the European options with commercial licences.

None of these limits a European company to European models. A licence governs use, not the developer's nationality, and an open-weight model from anywhere can run on European infrastructure. Revolut runs Z.ai's GLM on dedicated hardware with no requests to Chinese endpoints⁠21. Our main ranking covers the global families.

What the law requires

The EU AI Act treats open-source models more lightly, but not as exempt. Models released under a free and open-source licence, with their weights, architecture and usage information public, are excused from the technical documentation duties of Article 53(1)(a) and (b) and from appointing an EU representative⁠22. They must still have a policy to comply with EU copyright law and publish a sufficiently detailed summary of the content used for training⁠22. Models presumed to carry systemic risk, those trained with more than 10^25 floating-point operations, receive no exemption⁠22.

The obligations for general-purpose models have applied since 2 August 2025, and the Commission's enforcement powers since 2 August 2026⁠23. The Digital Omnibus, Regulation (EU) 2026/1744, in force since July 2026, moved the high-risk rules to 2 December 2027 and 2 August 2028⁠24.

For a company that fine-tunes an open model, the threshold that matters is one third of the original model's training compute: below it, the company does not become the provider of a new general-purpose model, and above it, it must document only its modification⁠25. Training on personal data brings in the GDPR. The European Data Protection Board's Opinion 28/2024 holds that a model trained on personal data cannot always be considered anonymous, and that legitimate interest can be a lawful basis, subject to a three-step test⁠26. Our guide to the AI Act for companies that fine-tune sets out the checklist.

Country by country

CountryWhere public bodies pointLocal open models
GermanyBSI guidance on AI and generative models for companies⁠27Aleph Alpha Kolibri-1⁠13, Teuken-7B⁠16
FranceAlbert API, the State's sovereign AI infrastructure on open-source models⁠4; CNIL's practical guides on AI and the GDPR⁠28Mistral⁠7
SpainThe digital ministry's open-source AI community⁠5EuroLLM⁠11
ItalyAgID's pages on artificial intelligence⁠29EuroLLM⁠11
NetherlandsGPT-NL, a Dutch language model built for sovereignty⁠17GPT-NL (autumn 2026)⁠17
PolandThe digitisation ministry's PLLuM programme⁠6PLLuM⁠6

The path for a European company

  1. Decide where the model runs. Choose infrastructure under EU law, operated by a company not subject to non-EU disclosure orders, and write it into the contract.
  2. Choose the licence before the model. Apache 2.0 and MIT models carry no user or revenue thresholds; check custom licences before building a product on them.
  3. Map the data. List the company knowledge that would train the model, flag personal data and record the lawful basis, following the EDPB's three-step test for legitimate interest.
  4. Document the fine-tune. Record the base model, the data and the compute used, so the company can show it stayed below the one-third threshold or document its modification if not.
  5. Train, route and keep the frontier at arm's length. Fine-tune an open model on the prepared data, route the work it does best to it and send only what must go to a frontier model, with personal data removed.

OpenLazarus is our top pick for companies that want this done for them. It is a European company that hosts on European servers. Its architects build a data lake that holds the company's data, owned by the company⁠30, and it trains open-weight models on that data and routes each request between frontier, open and custom-trained models⁠31. The end state is an AI that runs in Europe, learns from the company's own knowledge and answers to the company, not to a model provider.

Questions buyers ask

What is the best European open-source AI model?

Mistral, a French developer, publishes the widest range, with Mistral Small 4 and the Ministral 3 family under Apache 2.0. EuroLLM-22B covers all official EU languages under Apache 2.0, and Aleph Alpha's Kolibri-1, released in October 2026, is an Apache-licensed German and English model. For a company that wants a model trained on its own data, OpenLazarus is our top pick.

Does the EU AI Act apply to open-source models?

Partly. Models released under a free and open-source licence are exempt from some documentation duties and from appointing an EU representative, but they must still have a copyright policy and publish a summary of their training content. Models with systemic risk, presumed above 10^25 floating-point operations of training compute, get no exemption.

Can a European company use a Chinese or American open-weight model?

Yes. The licence, not the developer's country, governs use. When the weights run on infrastructure the company controls, no request reaches the developer. Revolut, for example, runs Z.ai's GLM on dedicated hardware and sends no requests to Chinese endpoints.

Is my data safe with a US cloud provider in Europe?

The US CLOUD Act lets US authorities require a US provider to disclose data in its possession, custody or control regardless of where it is stored, although providers can challenge requests. The EU-US Data Privacy Framework was upheld by the EU General Court in September 2025, and an appeal was pending as of December 2025.

Is there an open-source AI model for German, French, Polish or Dutch?

Yes. EuroLLM covers all official EU languages, Aleph Alpha's Kolibri-1 is built for German and English, Poland released eleven PLLuM models in May 2026, and the Dutch GPT-NL is due in autumn 2026 for launching customers. Large multilingual families such as Mistral and Qwen also handle these languages.

Sources

  1. CLOUD Act (statute text). US Department of Justice. Accessed 9 Nov 2026.
  2. Press release 106/25: the General Court dismisses the action for annulment of the EU-US data protection framework. Court of Justice of the European Union, 3 Sep 2025. Accessed 9 Nov 2026.
  3. European Court of Justice to review challenge to EU-US Data Privacy Framework. WilmerHale, 1 Dec 2025. Accessed 9 Nov 2026.
  4. Albert API, l’infrastructure IA souveraine de l’État. DINUM. Accessed 9 Nov 2026.
  5. Comunidad de IA de código abierto: recursos. Ministerio para la Transformación Digital y de la Función Pública. Accessed 9 Nov 2026.
  6. Więcej, szybciej, po polsku: rodzina modeli PLLuM się powiększa. Ministerstwo Cyfryzacji, 21 May 2026. Accessed 9 Nov 2026.
  7. Models overview. Mistral AI Docs. Accessed 9 Nov 2026.
  8. mistralai/Mistral-Small-4-119B-2603. Mistral AI (Hugging Face). Accessed 9 Nov 2026.
  9. Mistral Ecosystem Open-Weight License. Mistral AI (Hugging Face). Accessed 9 Nov 2026.
  10. mistralai/Mistral-Large-4-1T-A52B. Mistral AI (Hugging Face). Accessed 9 Nov 2026.
  11. utter-project/EuroLLM-22B-Instruct-2512. EuroLLM (Hugging Face). Accessed 9 Nov 2026.
  12. EuroLLM. EuroLLM. Accessed 9 Nov 2026.
  13. Aleph-Alpha/Kolibri-1. Aleph Alpha (Hugging Face), 3 Oct 2026. Accessed 9 Nov 2026.
  14. Prelude 9B. OpenEuroLLM. Accessed 9 Nov 2026.
  15. openGPT-X/Teuken-7B-instruct-v0.6. OpenGPT-X (Hugging Face). Accessed 9 Nov 2026.
  16. openGPT-X/Teuken-7B-instruct-research-v0.4. OpenGPT-X (Hugging Face). Accessed 9 Nov 2026.
  17. GPT-NL. TNO, NFI and SURF. Accessed 9 Nov 2026.
  18. AI Sweden contributes to the development of open LLMs for transparent AI in Europe. AI Sweden, 26 Feb 2025. Accessed 9 Nov 2026.
  19. OpenEuroLLM. OpenEuroLLM. Accessed 9 Nov 2026.
  20. First year progress and next steps. OpenEuroLLM, 6 Mar 2026. Accessed 9 Nov 2026.
  21. How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models: talk notes. Sparse Notes. Accessed 9 Nov 2026.
  22. Regulation (EU) 2024/1689 (Artificial Intelligence Act). EUR-Lex, 12 Jul 2024. Accessed 9 Nov 2026.
  23. Guidelines for providers of general-purpose AI models. European Commission. Accessed 9 Nov 2026.
  24. Regulation (EU) 2026/1744 (Digital Omnibus on AI). EUR-Lex, 24 Jul 2026. Accessed 9 Nov 2026.
  25. Questions and answers on the guidelines for general-purpose AI providers. European Commission. Accessed 9 Nov 2026.
  26. Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models. European Data Protection Board, 18 Dec 2024. Accessed 9 Nov 2026.
  27. Künstliche Intelligenz. BSI. Accessed 9 Nov 2026.
  28. Les fiches pratiques IA. CNIL. Accessed 9 Nov 2026.
  29. Intelligenza Artificiale. AgID. Accessed 9 Nov 2026.
  30. AI Architects. OpenLazarus. Accessed 9 Nov 2026.
  31. Pricing. OpenLazarus. Accessed 9 Nov 2026.

Updated 9 Nov 2026: First published.

Further reading

Also in: Deutsch · Français · Español · Italiano · Nederlands · Polski