Open models · European GPUs · pay per use
Chat, reading scans, understanding pictures, search by meaning, speech to text and picture editing, through one OpenAI-compatible API. Behind it, a growing network of small GPU nodes in Europe, each owned by the people who host it.
Each key has its own models and limits, and a statement per model and per day.
Most AI is served from data centres that burn electricity twice: once to compute, once to cool. We put the same GPUs where their heat is wanted, in the hands of the people who own them.
Eight ideas, one network. Point at an oval or a card: the two light up together.
A GPU turns nearly all the electricity it uses into heat. In a data centre that heat is thrown out, and more energy is spent getting rid of it. Our nodes stand in homes and workplaces, where the same heat warms the rooms.
Not one building full of servers but a network of small GPU nodes, each in a different place, joined by one API. Your call goes to a node that is free; nothing depends on a single site.
The people who host a node own it, and every call it answers earns them a share. The network grows because it pays the people who carry it.
Models run on more than one node. If one is down or busy, the next takes the call. A model that is not running anywhere refuses at once, so your fallback can take over.
Open-weight models on our own nodes, no third party in between. We keep counts of tokens, pictures and minutes for your statement, never your prompts or the answers.
Built and run from Bulgaria, in the European Union, under European law. Prices and statements in euros, or in dollars if you prefer.
Each node runs on the electricity of the place it stands in, used on the spot instead of carried across the grid to a distant data centre.
A handful of cards each doing real work: no halls of servers kept spinning for a peak that may never come, no chillers, no diesel backup.
Each call names the model it wants. Running now is live: a model that is not running is refused at once, so your fallback takes over without waiting.
Our everyday model: answers, writes, follows instructions, calls your tools and reads pictures. It streams its answer as it goes and remembers what it has already read in a conversation, so long chats stay quick and the repeated part costs a fifth of the price.
A larger model for the hard questions: grading and judging answers, long analysis, careful reasoning. Text only. Runs at night, 02:00 to 09:00 Eastern European Time; by day a call for it is refused at once, so your fallback takes over.
Reads the text in photos and scans: paragraphs, tables, forms and handwriting, in many languages. Hands the text back in reading order, so a scanned contract or a receipt becomes text you can search and quote.
Looks at a picture and tells you what is in it, reads what is written on it and answers questions about it, or sums up a scanned page. Also a quick, cheap text model for short jobs like sorting and summarising.
Turns text into vectors, so you can find things by what they mean and not only by their words, in more than a hundred languages: a question in Bulgarian finds the answer written in English.
Give it a question and a list of candidates, and it hands them back in order of how well each one answers. The last step that makes a search feel right.
Turns pictures into vectors, so a picture can be found with a few words, or with another picture that looks like it.
Writes down what was said, with the time of each line, in 25 European languages. Good for calls, voice notes and recordings.
Changes a picture the way you describe it: swap the background, change the clothes, add or remove a thing, and keep the rest as it was. One picture in, one picture out.
Checked at 16:39 UTC, 3 Oct 2026.
Per use, no subscription. Only answered calls are charged. Agreed prices for your account replace these.
| Model | Inputper 1M tokens | Cached inputper 1M tokens | Outputper 1M tokens | Pictureeach | Audioper minute |
|---|---|---|---|---|---|
qwen3.8-27bChat and reasoning |
€0.262$0.294 | €0.053$0.0595 | €1.87$2.10 | · | · |
qwen3-32bChat and reasoning |
€0.0356$0.04 | as input | €0.125$0.14 | · | · |
paddleocr-vlText from pictures and scans |
€0.0891$0.10 | as input | €0.0891$0.10 | · | · |
qwen3-vl-8bUnderstands pictures |
€0.0521$0.0585 | as input | €0.203$0.2275 | · | · |
embed-textSearch by meaning |
€0.00445$0.005 | · | · | · | · |
rerankSharper search results |
€0.00445$0.005 | · | · | · | · |
embed-imagePicture search |
· | · | · | €0.0000445$0.00005 | · |
parakeet-v3Speech to text |
· | · | · | · | €0.000668$0.00075 |
qwen-image-editEdit pictures with words |
· | · | · | €0.0134$0.015 | · |
Tell us what you are building and which models you need. You get a key with its own limits, and a statement every month.
Hostgarden