GPT-NL: A Sovereign Language Model for the Netherlands

GPT-NL establishes Dutch digital autonomy through a sovereign AI ecosystem

GPT-NL is an independent Dutch language model and ecosystem developed by TNO in collaboration with SURF and the Netherlands Forensic Institute (NFI). The project aims to reduce dependency on non-European AI providers by creating a model built on strong governance, transparency, and a commitment to public values, ensuring that the Netherlands maintains control over the technology, data, and ethical choices governing its AI infrastructure.

Core Principles and Governance

GPT-NL is designed around four primary values to ensure the model remains trustworthy and aligned with societal goals:

Sovereignty and Control

Developed within the Netherlands and Europe, GPT-NL provides full control over the model and its training data. This approach is intended to align the AI ecosystem with European laws and societal goals, avoiding reliance on external providers.

Transparency and Openness

The project emphasizes insight from source to model through several mechanisms:

  • Open Source Code: The source code is published as open source.
  • Dataset Insights: Detailed insights into the training datasets are shared publicly.
  • Controlled Licensing: Model weights are available under a controlled license to track usage and manage updates or data opt-outs.
  • Documentation: All choices regarding data collection, training, and the mitigation of bias and ethical risks are clearly documented.

Trustworthiness and Data Integrity

To eliminate risks associated with unclear data provenance or copyright infringement, GPT-NL is trained entirely from scratch. The data collection process follows strict criteria:

  • Anonymization and removal of personal data prior to training.
  • Exclusion of confidential or harmful content.
  • Safeguarding of intellectual property.
  • Prevention of data duplication within the dataset.

Reciprocity and Fair Value

GPT-NL implements a lawful data supply chain where data providers and rights holders are actively involved via a Content Board. This model ensures that a portion of the revenues flows back to the creators, shifting from a value-extraction model to a value-sharing one.

Resource Efficiency and Public Funding

Environmental Sustainability

Recognizing the high energy and water consumption associated with LLM development, the project focuses on optimizing model size and training processes based on scientific research to minimize its ecological footprint.

Financial Backing

The project is publicly funded by the Netherlands Enterprise Agency (RVO) on behalf of the Ministry of Economic Affairs and Climate Policy, with a total allocation of €13.5 million.

Technical and Strategic Critique

While the project emphasizes sovereignty, it has faced skepticism from the technical community regarding its feasibility and strategic direction. Key points of discussion include:

Budget and Capability Concerns

Critics have questioned whether a €13.5 million budget is sufficient to produce a competitive-quality model from scratch, with some suggesting the funding level might only achieve capabilities comparable to older models like GPT-2.

Sovereign vs. Open-Source Strategy

Some observers argue that true sovereignty should focus on the infrastructure (where compute happens) rather than the model architecture. Suggestions include:

  • Building on top of existing high-performance open-weight baselines (e.g., Qwen or Kimi) and fine-tuning them for specific agentic utility.
  • Collaborating at a European level rather than pursuing individual national efforts to achieve the scale required by scaling laws.

Economic and Regulatory Barriers

Discussion highlights a perceived gap between European regulatory goals (such as the EU AI Act) and the economic reality of AI development. Some argue that Europe lacks the venture capital and "urgency" seen in the US and China, which may hinder the ability of sovereign models to compete with frontier models.

"I keep seeing these 'sovereign' LMs time and time again... Instead of burning money on 'sovereign' claims, national research labs should instead focus on building on top of solid baselines... and finetuning frontier models with real agentic utility."

"The revenue sharing model is (IMO) more interesting than the fact that it is dutch. Usage of data, and sharing it with the providers of this data, is an inherent part of the creation of these models that is not discussed as much as it should be."

Sources