The Netherlands is moving ahead with its own national AI model instead of waiting for large foreign tech companies to solve the country’s language and data needs. Since late February, five organisations have been testing the beta version of GPT-NL through feasibility studies. The model was developed by TNO, the Netherlands Forensic Institute, and SURF, the IT cooperative for Dutch education and research institutions.
That group is expected to expand to 10 by spring 2026, with a wider commercial launch planned for the second half of the year.
The project has made notable progress since its launch in late 2023. At the time, it was met with skepticism because its budget was tiny compared with the spending power of major AI players. The key question was whether a public-interest model with limited resources could become something organisations would actually use, rather than just an experiment. A progress update published in February 2026 suggests the project is getting closer to that point.
From pre-training to practical use
GPT-NL has now completed pre-training. In the progress report, GPT-NL R&D manager Frank Brinkkemper said early benchmark results are promising. On summarisation tasks, the model already performs better than older systems such as GPT-3, which is notable given the gap in available resources. The team is using standard evaluation sets such as EuroEval, with adjustments where benchmarks are not fully tuned for Dutch.
The current feasibility studies are meant to answer the more important question: does the model work well in real environments for real tasks? Each customer works with a TNO team that installs the model on-premise, runs tests over three to six months, and then refines it for specific use cases. The customers pay for the research programme, including staff time, licensing, and an optional community fee. Final pricing for the wider market has not been set yet.
The first five participants are mostly public-sector users rather than private companies. Three of the use cases are funded by the Dutch Ministry of the Interior and Kingdom Relations, while TNO and the NFI make up the other two. That setup is practical, but the next stage will be more important for proving broader relevance across sectors.
The pilot projects cover several government and public-service functions. One involves Gem, a virtual municipal assistant already used by nearly 30 Dutch municipalities and responsible for close to 70,000 conversations in 2024. The project is testing whether GPT-NL can improve the quality of Gem’s answers to citizen questions compared with the models currently in use.
Another use case focuses on HIP, a communication assistant whose name roughly translates to Clear, Intelligent and Productive. The tool helps civil servants write government letters in plain language, which remains a major issue in Dutch public administration, especially for correspondence related to debt and benefits. GPT-NL is being compared with the commercial model currently deployed.
At the NFI, the model is being fine-tuned on forensic data to improve classification in investigations that involve terabytes of evidence, where both speed and accuracy matter for criminal proceedings. TNO is also using GPT-NL internally for classified and privacy-sensitive research under a “Copilot, unless” policy, meaning commercial AI tools are used by default unless data sensitivity requires a different option.
A licensing deal that could set a precedent
One of the most significant developments over the past year is legal rather than technical. GPT-NL has secured a licensing agreement with NDP, the Dutch association of commercial news publishers. The agreement covers national newspapers, platforms such as NU.nl and RTL News, and broadcaster BNR. GPT-NL says it is the first AI initiative in the world to reach paid, agreed licensing deals with all major publishers in a single market for training data use.
The deal was not simple to negotiate. News organisations have been among the hardest hit in the large language model era because their content has often been scraped without permission and then used in systems that can compete with journalism itself.
According to Rien van Beemen, chair of NDP Nieuwsmedia, the agreement creates a long-term precedent that strengthens journalism in the Netherlands. He said AI innovation can move forward ethically without the large-scale unlawful use of journalists’ work.
To address publisher concerns about content reappearing from the model, GPT-NL has added technical safeguards designed to reduce the risk of licensed material being extracted through prompts. The licensing terms are publicly documented, and the framework allows content to be withdrawn. At the same time, the team says it cannot continuously retrain the model on demand. Instead, providers who leave the project continue to receive compensation until a new model version is released.
According to GPT-NL product manager Saskia Lensink, the project has also drawn attention outside the Netherlands. Other European member states are asking how the legal and organisational structure works, and the project says the Netherlands is the first country to reach a collective agreement with publishers.
The case for digital sovereignty
GPT-NL is also built around a broader argument about digital independence. European governments, public institutions, and businesses depend heavily on non-European cloud platforms, office software, and AI systems. GPT-NL is rooted in non-profit, public-sector organisations such as TNO, SURF, and the NFI, which Lensink points out will never turn into American companies.
Lokke Moerel, professor of Global ICT Law at Tilburg University and a partner at Morrison Foerster, who advised the project on its legal structure, says the effort is strategically important. If Europe only writes rules for technology built elsewhere, it will always be reacting after the fact. In her view, building technology locally is the only way to avoid long-term dependence on foreign suppliers and to maintain bargaining power.
She also says the project is still at an early stage. The real challenge now is turning it into something scalable that becomes embedded in society.
Moerel says the Netherlands is often good at experimenting but less consistent when it comes to sustained investment. Her view is that this is the moment to push forward.
That gap between ambition and resources is clear. A team of about 25 people across several organisations has built a model that performs reasonably well on standard benchmarks, but scaling it into something that can complement or compete with frontier models will require far more funding than the €13.5 million public budget can cover. The model weights are not fully open source; they are available on request, with costs designed to recover ongoing operating expenses under the subsidy rules.
The team is aiming for a wider rollout in the second half of 2026 through professional licensing, with a hosted SaaS option also in development. It is also exploring a next-generation model with multilingual support and better speech capabilities. In the shorter term, improved retrieval-augmented generation functionality is on the roadmap.
The training dataset for GPT-NL v1.0 is expected to be published on HuggingFace, along with metadata covering the copyrighted content used, so the composition of the training data is transparent. Separately, a broad coalition including NDP Nieuwsmedia, TNO, VNO-NCW, and the National Library has called on the Dutch cabinet to create a formal data policy vision to speed up the adoption of responsible AI.
Whether a small team and a modest budget can make a lasting case will depend on what the next round of pilots delivers. The publisher agreement has already shown that the Netherlands can achieve something no one else has managed. The next test is whether the model itself is strong enough for organisations to choose it because it performs well, not just because it represents a principled alternative.
