Three years after OpenAI rolled out ChatGPT and brought artificial intelligence (AI) to the mainstream, several European countries are creating their own sovereign systems.
Sovereign AI is the ability of a country to develop, host, deploy, and manage homegrown AI systems for its citizens, rather than relying on foreign systems or cloud jurisdictions.
In a June report, the European Parliament acknowledged that the European Union has not been able to have its own technology champions because it “currently relies heavily on foreign technology, especially American technology.” The report said Europe’s dependence “looks set to continue” thanks to the US’ recent $500 billion (€432.9 billion) domestic AI. investment.
The EU said it needed to invest in research and develop new systems to regain its advantage. This is where governments can play an active role.
A small number of companies across Europe are building their own sovereign AI systems. Euronews Next takes a look at what has been built so far.
Germany
Germany is the latest country to announce its own AI plan, called Sovereign Open Source Foundation Models (SOOFI).
According to the German government, SOOFI is an attempt to create a basic “advanced AI” open source model that other companies developing AI products can adapt.
According to a statement from the German government, the technology will be used for highly complex tasks such as AI-controlled robots.
“With SOOFI, we are laying the foundations for a next generation European AI model that is sovereign, powerful and completely in European hands,” said Professor Wolfgang Nezil from Leibniz University Hannover, one of the universities participating in the project.
“AI models at scale that respect European values are essential to building trust in AI, especially in sensitive areas such as education, healthcare, public administration and production,” he added.
Telecommunications companies Deutsche Telekom and T-Systems said the goal is for SOOFI to have 100 billion parameters, or settings, that control the behavior of the model.
Both companies provide technical support for large-scale language models in one of their AI factories. To train the model, Deutsche Telekom will use around 130 NVIDIA chips and more than 1,000 graphics processing units (GPUs) that will be deployed by March next year.
Darmstadt University of Technology, one of the German universities participating in the project, said SOOFI will determine what is needed to build expertise in all areas of large-scale AI model development, from data collection and preparation to software construction and training.
Switzerland
In September, Switzerland’s AI initiative launched Apertus, the country’s first multilingual language model.
Apertus, which means “open” in Latin, allows researchers, professionals, and the public to customize the model to suit their specific needs.
The developers say that everything about the model is publicly available, including the training architecture, dataset, source code and model weights, and parameters that tell LLM how to interpret the data.
ETH Zurich, one of its collaborating universities, said Apertus was trained on 15 trillion tokens and pieces of information across more than 1,000 languages, including Swiss German and Romansh.
Apertus has been uploaded to Public AI, an online access point for sovereign models, so people around the world can access the model.
The Swiss AI Initiative said it would consider sector-specific models with expertise in law, climate, health and education.
“This release is not the final step, but the beginning,” Antoine Beausselet, co-leader of the Swiss AI Initiative, said of Public AI. “We are building for a long-term commitment to a sovereign, open AI foundation that serves the public good around the world.”
Poland
In February, Poland announced its national large-scale language model, the Polish Large Language Model (PLLuM).
PLLuM is “tailored to the specificities of the Polish language,” so AI speech and writing projects can “address the challenges of inflection and complex syntax well,” a government release said at the time of its launch.
The government suggests that PLLuM models can be transformed into AI to help write texts and emails, summarize documents, help students prepare for lessons, generate chatbot content, plan trips, and create summaries.
Poland’s Deputy Minister for Digital, Dariusz Standerski, said at the time that PLLuM was “an investment in a digital nation.”
At the time of the model’s announcement, Standerski said it would be extended to Hive AI. The system will eventually be integrated into the government’s administrative operations and help develop a “national AI ecosystem.”
For example, the public will have access to virtual assistants that help them obtain public information, and “intelligent” office assistants that automate document processing and information retrieval.
In the future, PLLuM will also be used to help teachers “deliver engaging lessons” using the latest technology in the classroom.
Spain
In January, the Barcelona Supercomputing Center (BSC) launched Alia, “Europe’s first open multilingual infrastructure” to develop “responsible AI for the service of people.”
BSC developed Alia with the help of the MareNostrum 5 supercomputer, which can perform 314 quintillion calculations per second.
Alia provides an open database of resources such as Spanish, Basque, Catalan, and Galician datasets, language models, and integration tools to help startups build their own national models.
The Spanish Artificial Intelligence Supervisory Authority (AESIA) has finally announced that Alia will be developed as a chatbot for tax authorities and integrated into an application that can easily diagnose heart failure.
The Alia project also builds on Ilena, another Spanish government initiative that has built more than 100 AI resources in Spanish, Basque, Catalan and Galician for use by domestic companies.
In 2020, the Catalan government will launched Aina is a pilot project that generates computer models in Catalan for other companies who want to develop AI products such as voice assistants, automatic translation, and conversational AI agents.
The government trained its model on an earlier Catalan database in which 1.7 million words were combined into 95 million sentences.
Netherlands
In 2023, three nonprofit organizations began developing a Dutch-speaking open source AI model called GPT-NL.
website DedicatedThe project describes GPT-NL as a model for “Dutch language and culture: authenticity, transparency, reciprocity and sovereignty.”
The consortium is usingIt is a mix of data obtained from high-quality sources through copyright agreements, publicly available data, and the generation of proprietary synthetic data.
The consortium recently signed an agreement with Dutch publisher and news agency ANP, part of trade association NDP Nieuwsmedia, to use articles for GPT-NL training. In return, the publisher receives a portion of the LLM’s profits when the LLM is finally released.
The project is also open source, meaning academic institutions, researchers and governments can try out applications in the health, education and service sectors. Users who are not using the LLM for professional reasons may have to pay a small fee to access the LLM once it is available.
The researchers hope to begin training the model in June 2025 and have the first version available. in front According to recent updates, the end of the year.
Portugal
Since 2024, a consortium of Portuguese universities has been working on developing a sovereign AI called Amalia.
Nova School of Science and Technology, one of the research groups behind Amalia, said it can answer questions, generate code, explain concepts, summarize text and interpret information in Portuguese and in the local context.
So far, researchers have been testing a beta version of Amalia in September and are working towards making the AI publicly available in mid-2026.
The government is already planning to use this large-scale language model to support analysis in government services and science through online portals.
According to local reports, Amalia will not be released to the public after being developed as a chatbot, but LLM’s code will be open source so other Portuguese companies can use it for their own AI models.