HostYourAI on Hugging Face
ProfileThis profile contains every model and dataset published by HostYourAI.
Why Loes is different
You can see what Loes is made of.
Covered by EU rules
Loes is trained and hosted on GPUs in the EU. We check the licence of each dataset and publish the data we use.
Bias is easier to trace
Because the training data is public, you can investigate where a skewed answer may have come from. If you spot one, we can trace it together.
Built with open material
The base models come from European labs and other open projects. Our own training data is public too. If you can improve something, we want you to get involved.
How it works
How we build Loes.
This is what happens between choosing a base model and putting a new version in the chat. We include the settings so you can check our work.
-
01
Use the best open-source models as a base
We continuously monitor new models. We use our own tests to decide which models are good enough to build on. Then we test them with exactly the same questions.
-
02
Assemble the training data
We combine Aya, xP3x, Dolly, and other public datasets. We check the licence of every set. Pirated books and web scrapes with unclear origins are left out.
-
03
Add Dutch conversations
Good public conversation data in Dutch is scarce. We therefore create some of it ourselves with other models, including models from Mistral.
-
04
Adapt the models
We adapt them to our European values, remove censorship, and make the Dutch sound more natural.
-
05
Store the weights
After training, we make our models public.
-
06
Test the Dutch
Every candidate receives the same held-out Dutch test. We consider its result alongside our practical test and standard benchmarks; we do not combine those different measurements into a single figure.
-
07
Make the model available
The chosen model runs with vLLM on European GPUs. The API is available through HostYourAI.
Models and data
It is all on Hugging Face.
We publish the models and training data behind Loes on Hugging Face. You can inspect or download the files yourself.
Loes
ModelOur Dutch fine-tune of Qwen3.8-27B. This is the complete merged model, approximately 55.6 GB.
Super Loes
2 modelsSuper Loes uses DeepSeek V4 Pro for text and Kimi K3 for questions with an image. Both model pages are linked below.
loes-xl-52k
DatasetThe dataset behind Loes, with 51,717 Dutch instruction pairs and the terms that apply to each source.
For developers
Loes in your own app.
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.
- Fully OpenAI-compatible, works with the SDKs you already use.
- In the API, Loes is hyai/loes-large and Super Loes is hyai/super-loes. Those names stay the same when we replace the model behind them.
- Processing stays in the EU and we never use your API traffic as training data.
- You pay for what you use, with no minimum charge. Streaming is supported.
The API uses the same European inference router as chat.loes.ai.
Terminal
curl 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
}'Measurement
What the numbers do and do not show.
HostYourAI measures quality with a weekly practical test. We measure speed separately as a technical snapshot.
27 scenarios · 8 categories
Our weekly practical test
Each selected model receives 27 fixed tasks through the same Loes route used by the chat, with our Dutch language layer in front of it. The test covers reliability, Dutch, sources, code, arithmetic, formatting, robustness and self-knowledge. The score is the percentage of usable scenarios that pass.
Latest verification · milliseconds
Speed remains separate
The speed ranking uses the complete response time of a short test during the latest technical verification on our EU infrastructure. Lower is faster. This is not a quality score: load and a cold start can affect the result.
The Loes models
Two models.
In the chat, you can choose between Loes and Super Loes.
Everyday use
Loes
Fast, detailed, and good in Dutch. Suitable for almost anything.
Difficult questions
Super Loes
Our strongest model. It reasons before answering, so it takes longer. Choose this for difficult questions. Still in beta.
Your conversations are never used to train models.
Why Loes
We show you what goes into Loes.
The data behind Loes is public, so you can look for blind spots and bias yourself. We work in small stages and measure whether each one genuinely improves the model. Only then do we spend money on the next training run.
Phase 1
Done
Train, publish, and launch Loes in the chat.
We tested several open base models, trained Loes on our Dutch dataset, and published the model on Hugging Face. The chat now lets you choose between Loes and Super Loes.
Days of work · tens of euros of compute
Phase 2
Next
Put more Dutch into the base model.
The next step is pre-training on roughly 100 billion tokens of Dutch text, including the open GPT-NL corpus, FineWeb-2, and Wikipedia. Then we fine-tune and test again.
Days to weeks · low thousands of euros
Phase 3
Planned
A version for law and government.
After that, we want to train a Loes for Dutch law and public-sector work. The sources will include public case law, legislation, and EUR-Lex, material that receives little attention in general models.
After phases 1 and 2
01 Contribute
Would you like to work on Loes?
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.
Data
Do you have Dutch or European text or instruction data with a clear open licence? It could help us add subjects that are still missing.
Compute
GPU time, credits, and spare capacity allow us to train and test models more often.
Expertise
We are looking for people who know NLP, evaluation, or linguistics. Knowledge of a field Loes does not yet cover well is welcome too.
Funding & partners
You can fund the project or work with us on a specific use of European AI.
Donate
A small contribution helps keep the free version online. You choose the amount when you donate through Stripe.
Send us a message
Tell us briefly what you would like to contribute and how to reach you.
Support Loes
Want to support Loes?
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.
Free
€0
/month
For anyone who wants to try a free European alternative to ChatGPT.
- Chat with Loes
- 25 messages every 5 hours
- 3 Super Loes messages every 5 hours
- Hosted entirely in the EU
- We never train on your chats
Most popular
Pro
€18
/month
For people who use Loes often and do not want to wait for a new message allowance.
- Everything in Free
- Unlimited chatting
- Image generation with FLUX
- Faster, larger models
- Priority during busy hours
- Early access to new models
Your own Loes
Custom
Your own private Loes environment for the people in your organisation.
- Everything in Pro, without limits
- Your own environment on your own address
- Dedicated compute in the EU
- Data processing agreement and clear terms
- Your own branding and knowledge base
🇳🇱 Every plan runs in the EU. We never use your conversations to train models.
Dutch AI.
The models and data are public.
We train and host Loes ourselves on European GPUs. You can try her now or help us build the next version.