Frequently asked questions

Walnoot 8B is an open-weight Dutch language model from WAINUT. You download it and run it on your own hardware or in your own cloud environment. The weights, the evaluation logs and the training recipes are under Apache 2.0. The model was continued pretrained and fine-tuned on top of Apertus, the open European base model that the Swiss AI Initiative at ETH Zürich and EPFL released under Apache 2.0.

Walnoot 8B is made for the work that makes up the largest part of AI use in an organisation: answering questions about a supplied document, classifying and routing incoming post, extracting data from a text and drafting a standard text. Reading comprehension, named entity recognition and classification carry a measured score in the published evaluation. Routing on the document types of an organisation itself is intended use and has not been measured separately. A summary from the model is a first draft that you read over. For rewriting into plain language we advise against the model. Both limits are in the model card as well.

For Dutch organisations where restraint is an obligation. The first group is government: ministries, executive agencies, provinces, municipalities and public knowledge institutions. The second is regulated business, such as insurers, financial institutions, healthcare and energy. The third consists of closed environments, such as defence and critical infrastructure. What those groups share is a layered AI policy whose sensitive layer has no tooling today.

Walnoot 8B is built for well-defined tasks. For work that calls for a lot of general knowledge, broad context or long reasoning, larger models are a better fit. You can see that in the EuroEval results too. We publish every score, including the tasks where Walnoot performs less strongly.

• Open-ended reasoning and long, complex analysis
• Creative work without a brief
• Code generation
• Anything where a mistake has immediate serious consequences without a person in between

A frontier model is the better tool for that. Most organisations push their entire workload through the most expensive and most exposed option there is, while a large part of that work needs nothing of the kind.

Read the model card

Yes, that is what it is made for. The weights are under Apache 2.0 on Hugging Face. You run them on your own hardware or in your own cloud environment, on the ground you choose yourself. There is no mandatory call to an external API and no telemetry, so it also works fully disconnected. A model behind an API cannot do that. Continued pretraining on your own data can happen inside your own environment, without that data leaving it. What hardware you need and which system prompt belongs with the model are described on /proberen. How fast the model runs depends on your card and your memory.

The model costs nothing. The weights, the logs and the recipes are under Apache 2.0, permanent and including commercial use, without registration and without a form. You pay for your own computing power and there is no meter running per token. WAINUT earns from the work around it: advice, staff training, implementation and staffing of AI talent. How that works per step is on Getting started.

What you have downloaded of the 8B stays yours under Apache 2.0, including for commercial use. The usage policy of the base model, the Apertus LLM Acceptable Use Policy, applies as well. Anyone may build on Walnoot, including a competitor. Openness is our philosophy. Later, larger models may come with different commercial terms. We make that distinction clear from the start.

Hosting it yourself removes one objection. The assessment stays: the DPIA, the legal basis, the accuracy requirement and the human in the loop stay with the organisation that deploys the model. We supply the artefact that the assessment is about. An approval is not part of it.

WAINUT is the Dutch AI company behind Walnoot, founded by two people. It builds open components that Dutch organisations can run in their own environment. Walnoot 8B is the model and NL-GOV-MCP fetches the Dutch sources the model works on. WAINUT put 50+ AI projects into production and trained 1000+ professionals on the work floor, at Dutch government bodies and regulated companies. Walnoot 8B is built on what those organisations need and therefore sized to their everyday work. The model is the free part. Deciding what you do with AI and getting it into production is the work.

We cannot rule that out. The training data as a whole was not filtered for personal data. Still, a sensitive part was removed and the data of the last training step was checked afterwards. It contained no Dutch citizen service numbers or bank account numbers. That check is coarse and does not show that the data is clean. The model can produce data that can be traced to a person. Whoever deploys it is the data controller for that. Do not use the model to remove personal data from documents either.

We call Walnoot 8B open-weight. The weights, the training recipe, the evaluation settings and the raw evaluation logs are under Apache 2.0, including for commercial use. We do not publish the training data itself. Open source would therefore promise more than there is. What is open per component is on /openheid.

A win only counts as a win when our whole confidence interval lies above their figure. A loss only when it lies entirely below it. If their figure falls inside our interval, the result is a tie. That holds even on a task where we are ahead on points.

We measured our column with EuroEval 17.6.0 on the test split, in 10 iterations and with bf16 weights. The GPT-NL column is their own publication of December 2025 on EuroEval 15.16.0, interim scores of the base model, published as point scores without an interval for the figures we take over. The harness is two major versions apart. Across versions the dataset content is not guaranteed to be identical. That is why we state explicitly with every comparison which version was used.

GPT-NL publishes no weights, so nobody outside the consortium can run that model. We have therefore never measured it ourselves. This table sets our measured figure next to their published figure.

Across the AI sector, digital sovereignty comes up more and more often. For an organisation it means that the data stays inside and that it decides for itself where the model runs. With Walnoot 8B you can check four things yourself. You hold the artefact yourself, downloaded under Apache 2.0 and yours to keep. You choose the jurisdiction it runs in, on your own hardware or in your own cloud environment, on European soil or fully disconnected. Nothing has to leave your environment, because there is no mandatory call to an external API and we put no telemetry in it. And it is built in the Netherlands, on an open European base model. It is about this model; about the country we make no claim here. Where you run it is a different question from where it was built.

Yes. Apache 2.0, permanent, including commercial use. No registration, no revocable licence.

Yes, for our layer. Documented per source, with the legal basis and the processing steps included. The layer underneath is Apertus. That documentation belongs to the Swiss AI Initiative and sits with them.

Yes, for the evaluation. The harness version is pinned and the command is on the site. You can trace the training through the recipe, but you cannot redo it exactly, because the instruction examples and the post-training data are not public.

European law for the instruction training, the post-training and the measurements. Those ran on MareNostrum5 in Barcelona, the European public supercomputer, via EuroHPC. Where you run the model is then up to you.