Europe’s Open-Source AI Pioneers: 10 groups Shaping LLMs Under the EU AI Act
AI is a technology that can help European organisations transform their business. For this, organisations can choose from many Large Language Models (LLMs), for a wide variety of use cases. However, most of these so-called frontier models are offered by American or Chinese companies. This will pose certain risks. For instance, how are they complying with the EU AI Act that will be in effect as of August of 2025? Do they enable data sovereignty, i.e. can you run them on your own servers? And what about support for any of the 24 official languages that are spoken in the EU? Perhaps, for a moment, we should divert our attention from American and Chinese suppliers and instead focus on European alternatives. In this blog post we introduce 10 European organisations who are offering open-source LLMs.
Aleph Alpha
Aleph Alpha GmbH is a German company that has been pioneering GenAI since 2019. With headquarters in Heidelberg it raised $500 million in a funding round backed by Bosch, SAP and Hewlett Packard Enterprise (HPE). Aleph Alpha is focusing on critical enterprises - organisations like law firms, healthcare providers and banks. These are organisations that rely heavily on trustable, accurate information.
Their motto is “Sovereignty in The AI era”, as can be seen from their website. What this means, in their own words, is taking responsibility - assuring data security, technology transparency and result explainability for AIs. As such they are aligned with the goals of the EU AI Act. But what they are not telling you on their website is that the company actively tried to influence European politicians to prevent the EU AI Act from imposing sanctions to AI companies who are not conforming to the new regulations (source).
Aleph Alpha has built its own model called Luminous. They have put a lot of emphasis on combating hallucination and researching ways to make its systems more reliable. It is, however a proprietary model, and it is not easily accessible. Luminous supports five European languages: English, German, French, Italian and Spanish. The Dutch language is not supported.
Pharia 1 is their newest model (source), which is only optimised for German, French and Spanish. Contrary to Luminous this small 7B model is an open source model, publicly available under the Open Aleph license, allowing for non-commercial research and educational use. If you think your use case applies, you can try it out using the Transformers library from Hugging Face (see model card for instructions).
iGenius
iGenius from Italy is promoting itself as a company that specialised in AI products. Founded in 2016 and headquartered in Milan, it supports mission-critical use cases, requiring maximum data security, reliability, and accuracy for customers in highly regulated sectors, like financial services, government, and heavy industry.
Its focus on data security for regulated industries would normally imply AI Act compliance. They have an article on their website dedicated to the EU AI Act but it does not share any details of what iGenius is actually doing in order to comply. However, they do offer a tool called Unicorn, that supports companies in adopting AI safely. Unicorn is powered by their own Large Language Models, which they have built in a partnership with NVIDIA, using Grace Blackwell Superchips.
Italia is the smaller model, released in July 2024. As the name suggests it excels in Italian - but only in Italian. Colosseum 335B is much bigger, and its performance is almost on par with GPT 4o and Claude Claude 3.5 Sonnet - at least according to some benchmarks, like ARC Challenge and MMLU (source). Colosseum supports no less than 50 languages, including Dutch.
Both models are accessible via NIM for Developers - the extensive API Catalog from NVIDIA that follows the OpenAI API. You can try them by first registering as an NVIDIA Developer. After that you can use either Python, LangChain, Node or Shell to send them your prompts. See this link for Italia and this link for Colosseum.
Italia is also available via Hugging Face on an MIT license, see this model card. But in order to use it you first have to fill out a form. Which may turn out to be a lot easier than interacting with the model in the Italian language.
LightOn
LightOn was the first European GenAI startup that went public (Paris stock exchange, November 2024). Founded in 2016, it developed hardware components for high-performance computers, before steering the company towards building LLMs in 2020. Now they have a strong list of enterprise clients.
As a European publicly listed company, LightOn falls entirely under EU oversight, providing additional assurance regarding compliance. In addition, the company has a sharp focus on data sovereignty, by putting confidentiality and value creation at the heart of their solutions.
They offer a proprietary LLM called VLM-4, which supports English, German, Spanish, French, Italian, but no Dutch. Regarding open source, LightOn was one of the participants of the BigScience initiative, thus contributing to the world’s largest open multilingual language model - BLOOM. BLOOM is a 176B open-access model that supports no less than 59 languages. But BLOOM is not part of their offering.
Instead they are offering an open-source LLM called PAGnol, which is a much smaller LLM. PAGnol only supports the French language. So for the Dutch market this isn’t quite interesting, but if you want to try it out anyway, you can access PAGnol via LightOn’s Hugging Face page. You can choose between small (125M), medium (355M), large (773M) and extra large (1.5B).
Mistral AI
Mistral AI is a well-known French startup, headquartered in the center of Paris. It was founded in 2023 by former AI scientists from Deepmind and Meta. Soon after it was established it raised €105 from a group of investors led by Lightspeed. This spearheaded the development and launch of their first LLM: Mistral 7B. It was released under the Apache 2.0 license.
Additional investment rounds followed and by now Mistral AI have developed an extensive library of frontier LLMs. Their latest offerings of premier models include Mistral Large, their top-tier reasoning model and Pixtral Large - a multimodal model (both released November 2024) as well as Codestral, a cutting-edge language model for coding (released January 2025). These models support dozens of languages, including English, French, German, Spanish, Italian and Dutch.
You can try Mistral Large via their chat bot, called Le Chat, available on the Web or on your phone (Android and iOS). API access is possible via their own platform, called La Plateforme (through here, see documentation here). Via La Plateforme you can also try out their legacy models, like Mixtral - their Mixture of Experts model. These models are also available via HuggingFace, Ollama and NVIDIA NIM.
In the summer of 2024, the company was valued at $6.2 Billion (source) which is still one order of magnitude less than Anthropic (source). However, their prospect for growth is enormous, which may also be a reason why they are not profiling themselves as a European player, in other words little league.
Being an open source company it is favored in the EU AI Act, but then again this could have been the result of extensive lobbying to influence legislation, back in 2023, even in their early days of their existence as a company.
Nyonic
Nyonic is one of the new kids on the block. Founded in 2023 and based in Berlin, they are dedicated to pushing the boundaries of modern language models, by tailoring them to specific industrial verticals, “for Europe”, explicity.
On their LinkedIn profile page, they recommend themselves as “building trustworthy and high-powered foundation models with European values at the core”. Trustworthy AI and European values are also at the core of the EU AI Act, so you would expect minimal regulatory risk. They are even saying that, apart from attracting ex-OpenAI scientist, they hired German AI ethicists.
Nyonic has built wonton, a language models that is available under the Apache 2.0 license, as wonton-3B and wonton-7B. You cannot use this model via Hugging Face, but you can have an idea of it through reading their Technical Report (see paper). The model is available for research and testing, but you would have to fill out a first (here and here).
The information is still sketchy and it has been awfully quiet around Nyonic for a year now. Which in this industry is a very, very long time.
Occiglot
Occiglot is an informal research community from Germany, whose aim is to develop open source language models. Occiglot is derived from the word Occident - the West - and by this they mean Europe.
Researchers from TU Darmstadt, the Hessian Center for Artificial Intelligence (Hessian AI) and the German Research Center for Artificial Intelligence (DFKI) are working together to build true European AI language models that are academic, non-profit and open source-based. In order to realise this they have been granted significant amounts of computing time on the institutes’ AI supercomputers.
Most models that are offered by the organisations in this post are State of the Art (SOTA) models. But not Occiglot. It is based on Mistral-7B, which was superseded by subsequent models. And while Occiglots aim is to create a coherent language model system that supports all 24 official languages of the European Union (as well as other unofficial and regional languages), the initial release only supports 5 languages: English, French, German, Spanish and Italian. Currently they are training models for other languages, including Dutch, but it is unclear when we might expect a new model release (source, source).
The models are available under the Apache 2.0 license, and can be found on their HuggingFace and page.
Occiglot is not operating at breakneck speed, like its American or Chinese counterparts. One thing we do expect is that it would align to the principles of the EU AI Act. But it does not seem as a viable LLM alternative as of yet.
OpenEuroLLM
OpenEuroLLM is a large pan-European consortium of organisations. The consortium includes universities (for instance Eindhoven University of Technology), companies (among which Silo AI, Aleph Alpha and LightOn, who are also mentioned on this list) and HPC centers (e.g. SURF). With an operational budget of €37 Million brought in by the European Commission and lots of HPC (high performance computing) power, the consortium expects to be able to compete with American and Chinese AI initiatives.
EuroLLM targets public and research sectors, to offer “a series of foundation models for transparent AI in Europe”. As can be expected from an EU funded initiative the OpenEuroLLM consortium fully aligns with the AI Act, emphasising transparency and European values. They even received an EU award for it.
The pre-trained and instruct models are available in sizes 1.7B and 9B. You can find them on their Hugging Face page, and from there you can follow instructions to use them via the Transformers library.
The models support all 24 European languages, which of course also includes Dutch. Russian and Arabic are also supported.
EuroLLM—9B scores well on benchmarks, compared to models from Mistral, Occiglot, Aleph Alpha and OpenGPT-X - organisations that are also mentioned in this post.
OpenGPT-X
The name OpenGPT-X suggests ties with OpenAI and xAI. But other than the fact that it trains large scale AI language models it has nothing to do with these American counterparts. Instead it aims to drive innovative language application services for businesses in Europe, which is emphasised in their motto: “Multilingual. Open. European”.
The OpenGPT-X consortium (participated by Deutsche Telekom, Fraunhofer, Aleph Alpha, among others) is funded by the the German Federal Ministry for Economic Affairs and Climate Action, ensuring AI Act and GDPR compliance. It develops language models and applications such as chatbots and digital assistants on the European GAIA-X cloud infrastructure, with bias prevention and digital sovereignty in mind.
With Teuken-7B OpenGPT-X has released a large AI language model that is designed for European needs. It is an open source, open weights model, available under the Apache 2.0 license, and made accesible through Hugging Face. See for the Teuken-7B-instruct-commercial-v0.4 model card, that illustrates how to use this model using the Transformers library.
The model supports all 24 European languages, including Dutch.
Silo AI
Silo AI was founded in 2017 in Finland. It is one of the largest private AI Labs in Europe. Their LinkedIn profile says it is focused on developing advanced AI models and solutions, and for this they employ some 300 AI scientists, including 125+ PhDs. They are offering custom LLMs through a SaaS subscription and they are profitable.
Also, in a partnership with Aleph Alpha they were promoting sovereign AI for Europe (source). To be sovereign means you have full control over where you run your AI workloads and how you manage data and operate your infrastructure. However, in 2024 they were acquired by AMD, and they are now using AMD compute platforms for their industry solutions. That is why their current name is AMD Silo AI. It is unclear as to what this involvement means for sovereign AI.
Silo AI has released two Large Language Models: Viking and Poro.
Viking supports the Nordic languages - Finnish of course, but also Icelandic, Norwegian, Swedish and Danish. If you understand either one these languages you may want to try it out on the LumiOpen community page on Hugging Face. Viking is available under the Apache 2.0 license, in sizes 7B, 13B and 33B.
Poro is meant to support all European languages, but at the time of writing language support is limited to just English and Finnish. Poro 34B is available on Hugging Face also on the LumiOpen community page and like Viking its license model is Apache 2.0.
Stability AI
Stability AI, founded in 2019 and based in London (UK), is a well known player among creators. For instance Stable Diffusion is a leading AI model for generating professional-grade images. But it has also built its own open source LLM, which is called Stable LM (now in version 2).
Of course, the United Kingdom is not part of the EU, but Stability AI is still a European player in the AI arena, that strongly favors open source models. Having said that it is unclear where the company is headed to. Also there have been reports of financial issues and weak economic stability, which may be a reason to be cautious if you would like to base your business on this language model.
Luckily there is no obstacle in trying their model on HuggingFace. On their page you will find Stable LM in sizes 1.6B, 3B and 12B. While you’re at it you can also play with other Stability models, like Stable Diffusion.
As an alternative you can also download StableLM 2 via Ollama, see this model page.