Dutch AI, hosted and trained in the EU
Loes runs in the EU. We never use your conversations for training. How we build Loes / Contribute
AI on our own terms
When a new open-weight model comes out, it goes through our benchmark of 33 fixed tasks. The best model of the moment becomes the base for Loes. Later we will choose per question, depending on load, which model gives you the best answer.
We train that model further with our own Dutch datasets: open data and conversations we make ourselves. There is always a fully European version on EuroLLM too, from base model to training data. Everything runs on GPUs in the EU.
Baby Loes is our own base model, trained from scratch on mostly Dutch text. She is not as good as the open models yet, but she goes through the same benchmark. As soon as she is better, she becomes Loes.
The Loes models
We keep benchmarking the best open-weight models, fine-tune them with our Dutch datasets, and make sure you can use them in Loes or Super Loes. So the model behind them can change from week to week.
For most questions. Answers quickly and is trained on Dutch.
For harder questions. Thinks first, so it takes longer to answer. Still in beta.
Here you can find our published fine-tunes and dataset, along with the models used by Super Loes. Baby Loes is listed as an ongoing project and is not yet public.
Our profile with the models and datasets we have published.
Our fine-tune of EuroLLM-22B-2512, using European data and training in the EU. Our published versions are linked below. The API name is hyai/loes-large-eu-eurollm-22b-2512.
Our current fine-tune of the Chinese Qwen3.8-27B, focused on Dutch language use, tone and following instructions.
Loes grows along with the open models. She first ran on Qwen3-14B, then on Qwen3.5-27B. Each time the benchmark pointed to a better model, we trained her again. The old versions stay public.
Our own base model built from scratch, with her own tokenizer and mainly Dutch text. She has 1 billion parameters, with 150 million active per token. She is not yet public.
Super Loes currently runs on DeepSeek V4 Pro, and on Kimi K3 for questions with an image, with our own Dutch layer in front. We are preparing our own fine-tune, with training data we make with Mistral Small 4 and EuroLLM. The links go to the current models.
The dataset behind Loes on Qwen: around 11,717 conversations we made with Gemma, 20,000 examples from Aya and 20,000 from xP3x. The current training run used 51,283 examples.
From base model to Loes
Loes always uses the best open-weight models of the moment. We benchmark every new model, fine-tune the best one with our Dutch datasets, and offer it as Loes. This is how that works, step by step.
Every new open-weight model goes through our benchmark. The best model of the moment becomes the base for Loes. There is always a version on a European base model as well.
We train with open Dutch datasets and with conversations we make ourselves with open models. For the European version we only use data made with European models.
Good Dutch conversation data is scarce, so we make some of it ourselves. That teaches Loes to write natural Dutch and follow instructions well. We never train on conversations from users.
When a better model comes out of the benchmark, we fine-tune Loes on it. The fine-tune mainly changes her Dutch, her tone, and how she follows instructions. Knowledge and reasoning largely come from the base model.
The European version builds on the best European base model and trains only on European data, in Europe. So there is always a Loes that is European from base model to training data.
Every version goes through the same practical test. We also compared the European version with her base model using the official EuroEval test for Dutch.
We offer the best version as Loes, on GPUs in the EU. We publish the models and data on Hugging Face.
For developers
You can call Loes from your own software too. The API follows the OpenAI format and runs on HostYourAI GPUs in the EU. In existing code, you usually only change the URL, API key, and model name.
Loescurl https://hostyourai.com/api/v1/chat/completions \
-H "Authorization: Bearer hyai-xxxxxxxx" \
-H "Content-Type: application/json" \
-d '{
"model": "hyai/loes-large",
"messages": [
{ "role": "user", "content": "Wie was Aletta Jacobs?" }
],
"stream": false
}'For organisations
A European alternative to ChatGPT Enterprise, Microsoft Copilot and Gemini.
A separate Loes environment with your name, branding and agreements, built for the GDPR from the ground up. That lets staff work with personal data and confidential documents too: everything stays in your own environment in Europe, we act as processor under a data processing agreement and nothing is used to train models.
jouworganisatie.loes.ai
Pricing
You can continue to use Loes for free, but the European GPUs still cost us money. Pro helps pay for them and gives you unlimited chat, image generation, and faster models. Your conversations remain in the EU and we never use them as training data.
For anyone who wants to try a free European alternative to ChatGPT.
For people who use Loes often and do not want to wait for a new message allowance.
Your own private Loes environment for the people in your organisation.
What we are working on
These activities run alongside each other. We continue developing Loes on open base models while training Baby Loes from scratch. This lets us offer an alternative today while working towards full sovereignty.
We focus on open models and open technology to make AI possible in the Netherlands. At HostYourAI we build the sovereign infrastructure for it, and with Loes the tooling.
We are building a Dutch alternative. Loes runs on servers in Europe and learns from Dutch language and Dutch conversations. We start with the best open models and work on our own model, Baby Loes, at the same time. Step by step, Loes becomes truly ours.
Your conversations stay yours. We never use them for training.
Phase 1
AvailableWe have fine-tuned Loes on Qwen and Loes EU on EuroLLM. The models are published. Loes is available in the chat and Loes EU through the API.
Phase 2
OngoingWe test new open models as soon as they are released and fine-tune the best model for Loes with Dutch data and conversation data. Alongside this, we continue developing a fully European track. This lets us track how European models compare with the other candidates.
Phase 3
StartedBaby Loes is our own base model built from scratch and is currently training. Our benchmark determines which model we offer under the name Loes. Our own models will also be considered once they are good enough. Later, we want to select the model expected to give the best answer for each request, taking the current load on our servers into account.
Built by HostYourAI
HostYourAI builds European AI infrastructure: the router and the servers Loes runs on. From Groningen we work on Loes, on our own model Baby Loes and on AI for organisations.
Groningen, Nederland
Groningse Jantine (34) bouwt AI-chatbot
In Groningen wordt gebouwd aan Nederlandse AI: "we zijn te afhankelijk"
Jantine Doornbos bouwde een Nederlands alternatief voor ChatGPT en Claude: "Loes is een vriendin"
Nederlandse ontwikkelt Loes AI, alternatief voor ChatGPT
Loes AI: Nederlands alternatief voor ChatGPT, draait lokaal en veilig
"En daar is Loes, de Nederlandstalige AI-chatassistent"
Groningse ontwikkelt eerste Nederlandse AI-assistent Loes AI
"Dit is gewoon de Nederlandse ChatGPT"
Contribute
We could use help with data, compute, and subject knowledge. If you have something that suits Loes, send us a message. We read it ourselves.
Tell us briefly what you would like to contribute and how to reach you.
We continue testing new open models and improving Loes while training Baby Loes. You can use the chat or contribute data, computing resources or expertise. We never use your conversations for training.