Dutch AI Models: What’s Available and What Each One Does Best

Dutch is a smaller language in the AI world, and the big global models have never made it a priority. That gap has pushed universities, research institutes and independent developers to build their own models, ranging from a government-backed national language model to compact open models that run on a laptop. Some are general chatbots, others are specialised tools for reading medical records or understanding text. Here is a clear look at the main ones and where each fits.

Model	Developer	Type	Best known for
GPT-NL	TNO, SURF and NFI	Large language model	Lawfully sourced training data and GDPR compliance
GEITje	Edwin Rijgersberg	7B language model (Mistral-based)	Pioneering open Dutch LLM
BERTje	University of Groningen	Encoder (BERT)	Dutch text understanding tasks
Fietje	Bram Vanroy	2.7B small language model (Phi-2 based)	Fully open and efficient
Schaapje	Community release	2B chat model (IBM Granite based)	Conversation and RAG on light hardware
MedRoBERTa.nl	Dutch research groups	Encoder (RoBERTa)	Hospital and clinical text

GPT-NL

GPT-NL is the country’s flagship effort, built by TNO, SURF and the Netherlands Forensic Institute and trained on the Snellius national supercomputer. Public funding of roughly 13.5 million euros backed the project, and the first beta version reached launching partners in February 2026.

Features

• Trained on data that was lawfully obtained, including archives licensed from Dutch news publishers
• Aimed at tasks like summarising, simplifying text and extracting information
• Being piloted by government bodies, including a trial with the Gem virtual assistant for municipalities
• A broader rollout, possibly through hosting providers, is expected in the second half of 2026

Strengths

• Privacy First described it as the first LLM that demonstrably complies with the GDPR
• Strong fit for public sector work where data sovereignty and copyright matter
• Built around Dutch language and context rather than adapted from an English model

GEITje

GEITje was the model that showed what dedicated Dutch training could do. It continued the pretraining of Mistral 7B on a large amount of Dutch text, and a later conversational variant, GEITje 7B Ultra, added preference alignment to make it better at chat.

Features

• 7 billion parameters with a Mistral base
• Chat-tuned Ultra version trained on synthetic conversational data
• Spawned domain spin-offs such as FinGEITje for Dutch financial text

Strengths

• Raised the bar for Dutch fluency when it appeared
• Became the foundation many later Dutch projects built on and compared against
• Worth knowing about, though the original model is no longer available for download

BERTje

BERTje is not a chatbot. It is a Dutch BERT model from the University of Groningen, built to understand text rather than generate it. It was trained on about 2.4 billion tokens of varied Dutch material, which gives it a much broader footing than the multilingual BERT that relied mostly on Wikipedia.

Features

• Encoder-only architecture
• Trained on a large, diverse Dutch dataset
• Vocabulary updated in 2021

Strengths

• Reliable for classification, named entity recognition and sentiment analysis
• Small and fast compared with modern LLMs
• A long-standing academic baseline, so plenty of research and tooling exists around it

Fietje

Fietje takes Microsoft’s Phi-2 and adapts it to Dutch using about 28 billion tokens of filtered open web data. At 2.7 billion parameters it is small, yet benchmark results showed it competing with models roughly twice its size at the time of release. Instruction and chat variants are available, and the chat version is the one to use as an assistant.

Features

• Base, instruct and chat variants
• Quantised versions for low-resource machines
• Weights, datasets, training code and evaluation code all publicly released

Strengths

• Excellent for experimentation, teaching and reproducible research
• Runs on modest hardware without a GPU cluster
• Full transparency, which is rare among language models

Schaapje

Schaapje is a compact Dutch chat model built on IBM Granite 3.0 2B Instruct. It is designed for practical, everyday use rather than headline benchmarks.

Features

• 2 billion parameters
• Tuned for Dutch conversation and instruction following
• Works with retrieval-augmented generation setups

Strengths

• Light enough for local or on-premise deployment
• Good option for businesses that want a private Dutch assistant over their own documents
• Easy to test before committing to something larger

MedRoBERTa.nl

MedRoBERTa.nl is a specialist. It was pretrained on free text from real hospital records, and it is one of only two publicly available encoder models worldwide with that kind of training data.

Features

• RoBERTa-based encoder
• Pretrained on genuine clinical notes
• Targeted at medical natural language processing

Strengths

• Understands clinical shorthand and terminology that general models handle poorly
• Useful for coding, classification and information extraction from patient notes
• Shows the value of domain-specific pretraining over generic models

Choosing the Right Model

The right pick depends on the job. Government bodies and organisations with strict compliance needs should keep an eye on GPT-NL as it moves from beta into wider release. Developers who want a small, open model to build on will get the most from Fietje or Schaapje. For classic text analysis, BERTje remains a sturdy choice, and healthcare projects have a purpose-built option in MedRoBERTa.nl. GEITje is still the reference point for Dutch LLM history, even if newer models have taken over the practical work.