AI your team can use, on infrastructure you own.
Not every organisation needs a model built from scratch. Most need the thing everyone already knows how to use, running somewhere their material is allowed to go. That is what this is.
Quince AI sets up a private AI workspace on infrastructure you control: an open-weight model, an interface your team can use on the first day, and your own documents behind it. It runs on hardware you own, on a device at the edge, or in a data centre under Swedish jurisdiction. Scope is fixed and quoted after one short call, and you own everything we leave behind.
Four parts,
all of them yours.
The model
An open-weight model, sized to the work and to the hardware. Open weights matter for a reason that is not ideological: the version you approve stays the version you run, until you decide to move. Nobody changes its behaviour underneath you.
The interface
A workspace your team opens in a browser and understands immediately, because it works the way the tools they already use work. Accounts, access control, and an audit log that records who saw what.
Your own material
Your documents, contracts, manuals or case files, indexed so people can ask questions of them and get answers with the source attached. The indexing happens inside your perimeter. Nothing is uploaded anywhere to make it work.
The output
Finished Word, PowerPoint and Excel files rather than text to paste. Meeting recordings turned into written notes in Swedish. Where it earns its place, image generation and a code interpreter.
Three places,
one decision.
The choice is usually made for you by what is in the material. We would rather tell you which one applies than sell you the most expensive one.
| Hardware you own | Swedish data centre | On the device | |
|---|---|---|---|
| Fits when | The material must not leave the building, or you want the asset | You want no hardware to buy or maintain | The work happens where there is no reliable network |
| Up front | You buy a machine | Nothing | You buy the devices |
| Jurisdiction | Yours | Swedish, and we read the ownership before the contract | Yours |
| Works offline | Yes, and it can be air-gapped | No | Yes, by design |
| Who operates it | You, or us alongside you | Us, or a partner you choose | You |
| Rules this out | Nobody to keep a machine running | Material that may not sit with a third party at all | Anything needing the whole document archive at once |
Where a system sits on the Sovereignty Ladder is the same question asked another way, and that page carries the fuller comparison against a rented cloud API.
What you keep
- The hardware, if you bought it, and the model weights on it.
- The configuration and the documentation, written so another supplier could pick it up.
- A handover session, and the answer to what still runs if you stop paying us. It is: all of it.
- Your data, which never needed to leave to make any of this work.
What this is not
- Not a subscription to somebody else's model with a new logo on it.
- Not a chatbot for your website. This is for the people who work there.
- Not a research project. If the honest answer is that a hosted tool serves you fine, we will say so.
- Not a reason to climb higher than your material requires. Sovereignty costs money, and buying more of it than you need is waste.
One call, then
a fixed scope.
We do not quote before we know what is in the material and who has to use it, because those two answers change everything else. The call is short and costs nothing.
Before the call
Run the requirements check. Eight questions, three minutes, no email needed. It tells you the class of system your answers describe and what the law makes you log, and it means the call starts from something concrete instead of from a blank page.
After it
A written scope with a fixed price, what we install, what you buy, and how long it takes. If keeping it running afterwards is useful to you, that becomes ML Ops & improvement, and it is designed to end when your own people no longer need us.
What people ask
on the first call.
Is this as good as ChatGPT?
For questions about your own documents, drafting, summarising and translation, the gap is smaller than most people expect and keeps narrowing. For frontier reasoning and the newest multimodal work, it is not. The honest test is your own material, measured, before anyone commits. We would rather run that test than argue about it.
What does it cost?
Scope is fixed and quoted after one short call, because the two things that drive cost are what is in your material and how many people use it at once, and we cannot guess either. What we will not do is quote a number before we know, then revise it upward once the work starts.
We have no IT department. Can we still do this?
Yes, and it is the most common situation among the firms who need this most. It changes the design rather than ruling it out: fewer moving parts, more of the operation on us, and a plain answer about what happens when something breaks. If nobody at all can keep a machine running, a Swedish data centre is usually the better of the three options.
Does it work in Swedish?
Swedish text handling is good and speech to text on Swedish audio is measurably better than the general-purpose alternative. We will show you the numbers on your own recordings rather than asking you to take that on faith, because a general claim about multilingual support is not the same as a measurement on your material.
Can it produce an audit log we can actually use?
Yes, and for several sectors it is not optional. Systematic log follow-up is a legal duty under HSLF-FS 2016:40 for patient material, and access control you can evidence is what GDPR Article 32 asks for. Per-user logs you can read back are part of the build, not an add-on.
What happens if we stop working with you?
Everything keeps running. You hold the hardware, the weights, the configuration and the documentation, and the handover is written so another supplier could take over. An engagement that leaves you unable to leave has failed on its own terms.
Tell us what is in the material.
That one answer decides most of the rest. Run the requirements check first if you want the call to start from something concrete, or book straight in and we will work it out together.