Netherlands moves GPT-NL from lab to live: first pilots under way

From pre-training to practice

Pre-training of the model is complete. According to GPT-NL R&D manager Frank Brinkkemper in the progress report, early benchmark results are encouraging: on summarisation tasks, GPT-NL already outperforms older models such as GPT-3 – notable given the difference in resources. Standard benchmarks including EuroEval are being used to measure performance, with corrections applied where benchmarks are not fully calibrated for Dutch.

The feasibility studies now under-way are designed to answer the harder question: does the model hold up in practice, on real infrastructure, for real tasks? Each launching customer works with a TNO team that installs the model on-premise, runs a series of tests over three to six months and then iterates towards specific use cases. The customers pay for the research programme – covering staff hours, licensing and an optional community fee – while final pricing for the broader market has yet to be set.

The first five participants are, notably, not from the open market. Three use cases are financed by the Dutch Ministry of the Interior and Kingdom Relations, with TNO and the NFI as the other two. That proximity to the project’s own initiators is a pragmatic choice, but it does mean the claim of broad sectoral relevance will be tested more rigorously in the next phase.

The applications being tested span a range of public sector functions. One pilot involves Gem, a virtual municipal assistant already used by nearly 30 Dutch municipalities, which handled close to 70,000 conversations in 2024. The question being investigated is whether GPT-NL improves the quality of Gem’s responses to citizen queries compared to currently deployed models.

A second use case centres on HIP, a communication assistant whose name translates roughly as Clear, Intelligent and Productive. The tool helps civil servants draft government letters in plain language, which is a persistent challenge in Dutch public administration, where correspondence on debt and benefits can be notoriously hard for citizens to understand. GPT-NL is being benchmarked against the currently deployed commercial model.

At the NFI, the model is being fine-tuned on forensic data to improve classification performance in investigations involving terabytes of evidence, a domain where processing speed and precision have direct consequences for criminal proceedings. TNO is also testing GPT-NL internally for its own classified and privacy-sensitive research projects, under a “Copilot, unless” policy, meaning commercial AI tools are the default, but GPT-NL steps in where data sensitivity demands it.

A global first in copyright licensing

Perhaps the most internationally significant development in the past year is not technical but legal. GPT-NL has secured a licensing agreement with NDP, the Dutch association of commercial news publishers – covering national newspapers, platforms including NU.nl and RTL News, and broadcaster BNR. GPT-NL claims to be the first AI initiative anywhere in the world to have reached paid, consensual agreements with all major publishers in a single market for the use of their content in model training.

The deal did not come easily. News media have more to lose than most in the large language model (LLM) era: their content has been scraped at scale without permission, then used to generate outputs that compete directly with journalism.

“We set a precedent that strengthens the position of journalism in the Netherlands over the long term,” said Rien van Beemen, chair of NDP Nieuwsmedia, in the report. “AI innovation can happen ethically, without large-scale unlawful use of journalists’ work.”

To address publisher concerns about content re-emerging from the model, GPT-NL has built in technical measures to prevent licensed material from being extracted via prompting. All licensing terms are publicly documented, and the framework allows for content withdrawal – though the team is candid that continuous model retraining on demand is not technically feasible. Instead, data providers who exit the project continue receiving compensation until a new model version is released.

According to GPT-NL’s product manager Saskia Lensink, the approach has attracted attention from beyond the Netherlands: “Other European member states are asking how the ecosystem of public and private collaborating parties is legally and organisationally structured. We are the first country in the world to have succeeded in reaching an agreement with the publishers collectively.”

The sovereignty argument

Underpinning GPT-NL is a strategic argument about digital autonomy. Europe’s governments, public institutions and businesses run largely on non-European cloud platforms, office software and AI tools. GPT-NL is anchored in non-profit, public-sector organisations – TNO, SURF, NFI – which, as Lensink pointedly notes, will never become American companies.

Lokke Moerel, professor of Global ICT Law at Tilburg University and a partner at Morrison Foerster who has advised the project on legal architecture, frames it as a strategic necessity. “If you only make rules for technology that others build, you will always be chasing events,” she says in the report. “You need to build technology yourself, otherwise you remain dependent on foreign suppliers and have no negotiating position.”

Moerel is also realistic about where the project stands. “This is still a startup in some ways. The real challenge begins now: how do you make this something scalable, something that structurally becomes part of society?”

She points to a pattern she observes in the Netherlands more broadly: strong at experimenting, reluctant to follow through with sustained investment. “This is precisely the moment to push forward,” she adds.

That tension between ambition and resources is tangible. A team of around 25 people, spread across multiple organisations, has built something that now performs respectably on standard benchmarks, but scaling it to complement or compete with frontier models will require a level of continued investment that €13.5m cannot cover. The model’s weights are not fully open source: they are available on request, with costs structured to recover ongoing operational expenses as required by the subsidy conditions.

The second half of 2026 is the target for a broader roll-out via professional licensing, with a hosted software-as-a-service (SaaS) option also under development. The team is also exploring a next-generation model with multilingual capabilities and enhanced speech support. In the near term, improved retrieval-augmented generation (RAG) functionality is on the roadmap.

The training dataset for GPT-NL v1.0 is set to be published on HuggingFace, including metadata on copyrighted content used, to provide full transparency about the composition of the training data. Separately, a broad coalition including NDP Nieuwsmedia, TNO, VNO-NCW and the National Library has called on the Dutch cabinet to develop a formal vision on data policy to accelerate the adoption of responsible AI.

Whether a small team with a modest budget can make that case stick will depend on what the next wave of pilots delivers. The publisher deal has shown that the Netherlands can do something no one else has managed. Now the question is whether the model itself is good enough to make organisations choose it over the alternatives – not out of principle, but out of performance.