Why Loes is different
Sovereign by design.
Built for EU rules
Trained and hosted on GPUs inside the EU, under EU law, on a data mix you can audit. We build it to fit the rules from the start, not patch it up afterwards.
Bias you can inspect
Trained on data we can actually point to, in our own languages. So when it leans a certain way, we can show you why, and you can call it out.
Built in the open, together
Open base models from EU labs, clean licensed data, and room for anyone who wants to help push it forward.
How it works
Exactly how we build Loes.
No magic, and nothing hidden. Here's the whole pipeline, start to finish, with the real settings we use.
-
01
Line up two tracks
We always run two tracks side by side: a fully European base (EuroLLM-22B) and a strong counterweight on a different type of base model (Qwen3.5-27B). That gives you a fully European option and maximum quality, instead of betting on one horse.
-
02
Gather clean, licensed data
A hand-picked mix of openly licensed public data, Aya, xP3x, Dolly and more, with the licence checked on every single set. No pirated text, no scraped-from-nowhere corpora.
-
03
Fill the Dutch gaps ourselves
Good open Dutch conversation data is thin, so we generate our own, using an open European teacher model (Mistral), never a closed one like GPT-4. That keeps the result clean enough to actually sell.
-
04
Fine-tune with QLoRA
We train each base on that mix with QLoRA, LoRA rank 32, learning rate 2e-4, two epochs, 4k context, all on GPUs inside the EU. Light enough to do in hours, not weeks.
-
05
Merge and store the weights
The trained adapter is merged back into the base model, and the full weights are pushed to a private, EU-hosted repository. Nothing ever leaves European servers.
-
06
Test it on real Dutch
Every candidate gets the same held-out Dutch test. Only the one that clears the bar, a Dutch-quality score of 70% or higher, ships. The rest get dropped, no exceptions.
-
07
Serve it behind a normal API
The winner runs on EU GPUs with vLLM, behind a standard OpenAI-compatible API. Point your existing code at a new URL and you're done.
Everything in the open
The weights and the data are public.
No black box. The Loes models and the full training set live openly on our HuggingFace organisation, under clean licences. Download them, inspect them, run them yourself.
HostYourAI on HuggingFace
OrgOur open organisation, every model and dataset in one place.
huggingface.co/HostYourAI ↗Loes Large EU
ModelThe fully European option on EuroLLM-22B, quality score 97.5. Fully public and live in the chat.
huggingface.co/HostYourAI ↗Loes Large World
ModelThe strong counterweight on Qwen3.5-27B, fully public and the default in the chat since 10 July.
huggingface.co/HostYourAI ↗Loes EU
ModelThe first generation on the European track (EuroLLM-22B), merged weights, Apache-2.0.
huggingface.co/HostYourAI ↗Loes World
ModelThe first generation on the World track (Qwen3-14B), also fully public.
huggingface.co/HostYourAI ↗loes-xl-52k
DatasetThe full clean Dutch training set, 51,717 examples.
huggingface.co/HostYourAI ↗For developers
Loes in your own app.
Building software? Talk to Loes through one OpenAI-compatible API, hosted on GPUs inside the EU by HostYourAI. Point your existing code at a new URL, pick a Loes model, done.
- Fully OpenAI-compatible, works with the SDKs you already use.
- Pick your Loes model: the default is hyai/loes-large (World), fully European is hyai/loes-large-eu-eurollm-22b-2512; the first generation stays available.
- EU-hosted and GDPR-proof, we never train on your traffic.
- Pay per use, no minimum, streaming supported.
Powered by HostYourAI, the same EU inference router that runs chat.loes.ai.
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
}'The numbers, honestly
How good is Loes, really?
We measure everything ourselves: the quality score per Loes model, and the two tracks head-to-head on EuroEval-nl, the European standard benchmark. No marketing numbers, this is what we actually measured. And yes, honest about where Loes doesn't reach.
Internal measurement on Dutch prompts and grounding behaviour, 10 July 2026. The new Loes Large models (World on Qwen3.5-27B, EU on EuroLLM-22B) leave the first generation well behind; Loes Large World has been the default in the chat since 10 July. Per-model details under The Loes models below.
Scores 0-100 on EuroEval-nl (higher = better), measured 26 June 2026, per track. A different scale than the quality score above: this measures individual skills per category. The trade-off at a glance: the EU track wins on Dutch grammar (linguistic acceptability 43.5 vs 23.5), the World track is stronger on raw knowledge and reasoning (knowledge 72.9, common-sense 73.6). The Loes Large models build on exactly these two tracks; their EuroEval run is coming.
Where does Loes sit among European models?
We compare Loes against other European models, not against closed US models (which you can't self-host in the EU anyway). And here Loes stands strong: Loes Large EU noticeably lifts her European EuroLLM-22B base on Dutch, and now speaks it flawlessly. The only European model clearly scoring higher, the French Mistral-Small-24B, carries a non-commercial licence, so you can't use it in a product. So Loes wins as the strongest sovereign Dutch model you can actually use freely.
The Loes models
Two tracks, both trained fully in Europe.
Loes always runs on two tracks side by side: a fully European option on EuroLLM, and a counterweight on a different type of base model for extra quality and feel in the chat. Both get exactly the same, fully European training with open European teachers and our own data. Below are all four Loes models: the new Large generation and the first generation that preceded it. And the best part: the fully European option gives up nothing. EU and World are neck and neck, and EU even speaks slightly more flawless Dutch.
Loes Large World
Base: Qwen3.5-27B (dense)
The best all-rounder: strongest knowledge and reasoning. The default in the chat since 10 July.
- Format instructions
- 100%
- Dutch
- 92%
- Grounding + multi-turn
- 8/8
Loes Large EU
Base: EuroLLM-22B (European consortium)
Fully European from base to training, with the finest Dutch and the highest language score.
- Dutch
- 100%
- Format instructions
- 92%
- Grounding + multi-turn
- 8/8
Loes World
Base: Qwen3-14B
The default until 10 July, still selectable in the chat.
- Grounding + multi-turn
- 8/8
- Status
- Live in the chat
- Role
- Previous default
Loes EU
Base: EuroLLM-22B
The first European track, still selectable in the chat.
- Format instructions
- 67%
- Status
- Live in the chat
- Role
- European track
The quality score is an internal measurement on Dutch prompts and grounding behaviour, not an official benchmark. Higher is better, shown on a 0 to 100 scale. Measured on 10 July 2026. All models are open on Hugging Face under HostYourAI, and we never train on your conversations.
Why Loes
Transparent, with no hidden bias.
Why Loes: because you should be able to see what your AI was trained on. Loes runs on data we can actually point to, fully in the EU and built in the open. So you can spot bias instead of trusting it isn't there, and you know exactly what's underneath. We build her in stages, cheapest and fastest first, and each stage has to prove itself before we spend on the next.
Own finetune, open on Hugging Face, two tracks live.
The base bake-off, instruction-tuning on the clean mix and our own Dutch conversation data, all gated on a real Dutch eval. This one is standing: both Loes Large tracks are trained, measured, open on Hugging Face and live in the chat since 10 July, with Loes Large World as the default.
Days of work · tens of euros of compute
Close whatever Dutch gap is left.
We keep pre-training the base on around 100 billion tokens of clean Dutch text, including GPT-NL's open corpus, FineWeb-2 and Wikipedia, then fine-tune and test again. Pioneers like GPT-NL provide exactly the open building blocks a project like this can build on.
Days to weeks · low thousands of euros
Law and government, done properly.
A version that's genuinely good at Dutch law and public-sector work, trained on open sources the big frontier models barely touch, published case law, legislation, and EUR-Lex. This is where a sovereign model earns its keep.
After phases 1 and 2 land
01 Contribute
Help build Loes,
in any way.
No single team builds something like Loes. Every part of her is something someone out there could help with, and if that's you, we'd genuinely like to hear from you.
Data
Clean, openly licensed Dutch & EU text or instruction data, so we can cover more ground without wandering into legal grey areas.
Compute
GPUs, credits, or spare capacity, so we can train and test more models, sooner.
Expertise
NLP, evaluation, linguistics, or deep knowledge of a field, to help make Loes genuinely good.
Funding & partners
Back the project, or team up with us to push European AI forward.
Offer a hand
Tell us what you can bring. We read everything that comes in.
Support Loes
Want to support Loes?
Loes is, and stays, free to use. But EU GPUs, training and hosting cost money, and we do this without a US cloud behind us. Go Pro and you keep Loes independent and open, and get unlimited chat and faster replies in return. Your conversations stay in the EU, GDPR-proof, and are never used to train models.
Free
For everyone who wants a private, European alternative to ChatGPT.
- Chat with Loes
- 50 messages per day
- Fully EU-hosted, AVG/GDPR-proof
- We never train on your chats
Pro
For power users who want more, faster, without limits.
- Everything in Free
- Unlimited chatting
- Faster, larger models
- Priority during busy hours
- Early access to new models
Business
For teams and organisations with compliance needs.
- Everything in Pro
- Team management & invoicing
- DPA / data processing agreement
- SSO and private models
- Dedicated EU hosting
🇳🇱 Every plan is fully EU-hosted, AVG/GDPR-proof, and we never train on your conversations.
Made in the Netherlands.
Built in the open.
Trained and hosted in the EU, by the people who actually run the GPUs. Come and help us build it.