How Europe's network of open-source LLMs is evolving

By Ioanna Lykiardopoulou

In the global race for AI development, the EU has been following a distinctive approach: open-source Large Language Models (LLMs).  

The objective is clear: firstly, to create models that are transparent in their design and training data, aligning with the bloc’s commitment to trustworthy and responsible AI. Secondly, to release such models freely for anyone to build their AI applications and foster collaboration. And thirdly, to develop multilingual LLMs that serve the Union’s linguistic and cultural diversity. 

Examples of this approach include two EU co-funded initiatives: EuroLLM, which offers 9B and 1.7B parameter models, and OpenEuroLLM which is working on a series of foundation models. 

“Over the next year, we’ll start seeing the results of these projects,” says Edwin Rijgersberg, the creator of GEITje — the first open Dutch LLM — and founder of the AI Studio Delta, who spoke at TechDays this year. 

“But if you look at what’s been happening lately, it’s mostly about national initiatives,” he adds. These include France’s LUCIE, Spain’s ALIA models, and the Netherlands’ (albeit not fully open-source) GPT-NL. “Germany is also very active,” he says. 

Different paths, shared European values 

All these initiatives share key European principles: they’re (almost exclusively) open-source, transparent, and linguistically inclusive. Yet, they also demonstrate differences in both development and vision. 

Perhaps the most unique of the current national efforts is the Dutch GPT-NL project, according to Rijgersberg. Now nearing its completion, GPT-NL stands out with its commitment to using strictly opt-in data, released under permissive licenses, donated voluntarily, or provided through partnerships. 

“This is an interesting experiment and one that Europe can learn from. But it’s also a very challenging strategy to execute, especially for a language like Dutch that already doesn’t have much available text data to train on,” Rijgersberg says. 

That challenge doesn’t just affect training — it also shapes how open the model can be. Under its opt-in data approach, GPT-NL’s source code and openly licensed datasets will be made public, but the model weights and copyrighted material will remain behind research and professional licences.

Meanwhile, across Europe, countries are forming their own visions for AI. France, for instance, “is pouring a lot of money and attention into what I call ‘creating’ AI,” Rijgersberg says, “rejecting the idea that they have to rely on the US or China.” By contrast, the Netherlands’ focus is mainly on the correct and ethical use of generative AI applications, rather than on building the underlying GenAI models.

Still, Rijgersberg sees a critical benefit in this diversity of approaches: “All these country-level efforts gather data and experience that is shared all over Europe.”

“Countries don’t have to reinvent the wheel by themselves.” 

Towards a pan-European open-source network 

Performance-wise, European open-source LLMs do not yet match the best American and Chinese models, which Rijgersberg attributes to two primary reasons: the lack of computational power and the need to comply with the EU’s text and data mining rules. 

“If you look at benchmarks, with all these smaller initiatives, we have surpassed the performance of Meta’s Llama 2. I would say that we are approximately about a year behind, which is still a million years in AI.” 

But Rijgersberg remains optimistic, describing this as a “surmountable gap,” especially as the EU moves towards addressing the above two challenges. The AI Act now provides further clarity on data training requirements, while the bloc’s investment into 19 AI factories is expected to significantly increase computing power (including access to supercomputers), data resources, and technical expertise. 

The AI factories are also signaling a path towards a pan-European infrastructure, where open-source models have a big role to play. “We’re starting to see some cooperation and open-source is the best form of sharing,” Rijgersberg says. 

“It’s about putting everything in the open to allow people to learn from each other, come up with new ideas, and collaborate to create new pan-European initiatives.” 

The open-source approach offers benefits that go beyond the strategic imperative of digital sovereignty. It also ensures AI is accessible to everyone — from public institutions and major enterprises to small companies, universities, and individuals.

“The only way we are going to move our society forward in this space is by doing together," Rijgersberg remarks.

💡 Want to learn more about the KickstartAI community and our mission to accelerate AI adoption in the Netherlands? Check out our website, listen to our podcast, or follow us on LinkedIn!