Walnoot 8B: a Dutch language model you run yourself
Walnoot 8B is an open-weight language model by WAINUT. You download it and run it in your own environment, including when data is not allowed to leave that environment. The model costs nothing.
Open-weight · Apache 2.0 · Built by WAINUT
The model
What Walnoot 8B is
Walnoot 8B Instruct is a Dutch language model with 8.05 billion parameters. It is built on Apertus-8B-2509, the open European base model that the Swiss AI Initiative (ETH Zürich and EPFL) released under Apache 2.0. WAINUT continued training it on Dutch text and gave it its own Dutch instruction layer.
You download the weights and run them on your own hardware or in your own cloud environment. There is no mandatory call to an external API and no telemetry, so it also works fully disconnected (air-gapped). Walnoot 8B is available as a download only. There is no chat app, hosted API or waitlist.
8.05B
Parameters
Apertus
Base model
Apache 2.0
License
Measured with EuroEval 17.6.0 on Walnoot 8B Instruct 1.0
The original goal has been met: match GPT-NL on all six published benchmarks. Five times above it, level on dbrd.
The scoreboard on walnoot.ai shows all ten tasks with their confidence interval. It also lists the command to recompute every score yourself.
View the scoreboardWhat it is built for
Walnoot 8B is built for work that can be bounded.
Answering questions about a document
Ask a question about a supplied document and have the answer taken from that document.
Reading comprehension (squad-nl): 62.08 vs 51 for GPT-NL
Extracting information from a text
Pick names, places and organisations out of a Dutch text so you can carry them over into a form or a register.
Named entity recognition (conll-nl): 46.52 vs 36 for GPT-NL
Classifying texts
Label incoming post, tickets and documents and route them to the right place.
Measured through sentiment (dbrd): 90.56 vs 90 for GPT-NL, a tie
Drafting standard texts
Draft a letter or a memo to a fixed pattern, using the details you supply.
We have no measured figure for this
Code generation, open-ended reasoning and long, complex analysis are out of scope, as are applications where a mistake has immediate serious consequences without a person in between. For that work a larger model or a person is the better tool. Rewriting into plain language also falls outside the intended use. A summary from Walnoot 8B is a first draft that you check.
Who it is for
For Dutch organisations where restraint with data is an obligation.
Government and public organisations
Ministries, executive agencies, provinces and municipalities. The model runs in their own environment, with documented provenance for whoever has to assess it.
Regulated sectors
Insurers, financial institutions, healthcare and energy. The concern here is the sensitive part of the document flow, which has to stay within their own risk management.
Restricted environments
Defence and critical infrastructure. Walnoot 8B runs disconnected, on a fixed version with a hash that you can check again later.
What WAINUT does
From download to everyday use
Around the model, WAINUT helps in three ways.
Decide what goes where
With the AI Menu we work out which AI belongs to which task in your organisation, before you choose a model.
Build it in
We set up Walnoot 8B in your own environment and can continue training it on your own data, without that data leaving the environment. Where Dutch government sources matter, NL-GOV-MCP retrieves the documents and Walnoot 8B reads them.
Built by WAINUT
Walnoot is a brand of its own with its own site, built by WAINUT. We work every day with organisations that handle sensitive data. There we saw applications stall, because the information is not allowed to leave their own environment.
So we set one goal in advance: match GPT-NL on six comparable tasks. From there we built a compact Dutch language model for bounded tasks.
Frequently asked questions
Deploying Walnoot in your organisation?
Book a call. Together we look at which tasks suit it and what it takes to run the model in your own environment.