Løynd, private AI that stays in Norway

The problem it answers

There is a large category of Norwegian work that would obviously benefit from a language model and cannot use one, because the text in question is a patient note, a client file, a supplier invoice, or a contract that is not allowed to leave the country. The usual answer is a data processing agreement and a promise. Løynd's answer is that the prompt never goes anywhere: the model is ours, the hardware is ours, and it sits in Norway.

The name is Nynorsk for hidden, which is also the design constraint. Retention is not a setting buried in an admin panel that defaults to on. Every API key starts at storage mode none, and the opt-in modes encrypt with a key we do not keep, so an encrypted history is unreadable to us by construction rather than by policy.

How the topology enforces it

The privacy claim is only worth as much as the network path behind it, so the path is the interesting part of the build. The website is static assets on Cloudflare and nothing else. The chat widget in the browser posts straight to api.loynd.no, which is a DNS-only record pointing at the office server, so prompt data goes from the visitor's browser to Norway without transiting Cloudflare or any other third party on the way. Third-party requests are absent from the site entirely: no CDN fonts, no analytics, no cookies.

The database on the server stores API keys and aggregate token counts. It does not store prompt content, which keeps the retention promise a property of the schema rather than of a deletion job that has to keep running correctly.

What you can use

• A free Lite model, demoed straight on the front page with a streaming chat widget and no login.
• An OpenAI-compatible API: POST /v1/chat/completions with a bearer key, so existing clients and SDKs point at it by changing a base URL.
• Stronger models and API access on request, which is deliberate: capacity is physical machines in a room, not an elastic bill.
• Optional encrypted history, where the key is the user's and the ciphertext is all we hold.

Myntos is the first production consumer: it uses a small Løynd-hosted model to pick an accounting category off an uploaded invoice. That is the exact shape of the problem Løynd exists for, since the document is a real company's books and sending it to a foreign API is not an option.

Stack

• Backend: Python 3.12, FastAPI and Uvicorn, Hugging Face Transformers serving a local open-source model, SQLite from the standard library for keys and counters. Runs on our own server in Norway.
• Frontend: Astro, static output, no JavaScript framework and no third-party requests. Norwegian by default, English at /en/.
• Two independent repositories, because the halves deploy to different machines: the site to the edge, the API to the server that holds the model.

Explore