Quantum Howl

Local AI for your organization

Models, data and agents that run inside your own infrastructure (on-premise). Dental Brain has worked this way since January 2026 in a clinic with two sites.

DGX Spark
Slow-motion simulation. The model runs on your server; if any phase ever needed data to leave, it is agreed beforehand and in writing, with the data pseudonymised.

What local AI is (and what it is not)

Local AI is a model that runs on your server and works with the data you choose.

It is not a cloud chat, nor just installing a model on a computer. It includes the data the model works with, its evaluation against a metric agreed before we start, and maintenance afterwards.

How it differs from ChatGPT for business

The difference is where the model runs and who controls it.

CriterionChatGPT for businessLocal AI with us
Where it runsChatGPT for businessOn OpenAI's systems; your workspace admin decides how long conversations are kept. SourceLocal AI with usOn your server.
Who controls the model and its versionsChatGPT for businessOpenAI, which decides which models it offers and when it retires older ones. SourceLocal AI with usYou, with us: the version changes only when you agree.
What leaves your infrastructureChatGPT for businessWhat you send to the service. OpenAI states that, by default, it does not use your business data to train its models. SourceLocal AI with usOnly what we agree beforehand and in writing, with the data pseudonymised.
How it is paidChatGPT for businessPer user per month. SourceLocal AI with usFree feasibility study; the pilot is quoted on request, at a fixed price.
When it fitsChatGPT for businessIf your data has no special rules and you need it now, a cloud service is quicker to start.Local AI with usIf your data has special rules, or you want to decide which model you use and where it lives.

OpenAI information checked on 19 November 2026.

What hardware you need

It depends on the case. In the feasibility study we look at what you have and, before we start, tell you what would be needed and what it would cost.

Our reference: we serve models with vLLM on an NVIDIA DGX Spark. The model starts answering in 24.72 ms (June 2026, Sprint DGX).

Training at scale is a different matter: 60 days on a node of 8 NVIDIA H100 GPUs (2026, NVIDIA Innovation Lab grant). It is not something you have to install in your organization.

GDPR and the EU AI Act

Having the model on your server does not exempt you from the GDPR.

If we access the system to maintain it, we sign a data processing agreement with you (art. 28 GDPR).

For the EU AI Act we give you the technical documentation, as a basis for your legal advisor. We do not guarantee compliance: your advisor assesses it.

You decide which data the model uses and where the data is kept. We work on your infrastructure. If any phase needed data to leave it, for example to train on external GPUs, we would propose it beforehand and in writing, with the data pseudonymised, as we describe in chapter 3 of the Sprint DGX series.

Read the chapter

When it is not the right fit

Local AI is not the answer to everything. We will tell you no in these cases.

  • If your problem can be solved without AI or with a tool that already exists.
  • If there is not enough data, or it cannot legally be used for what you want.
  • If we cannot agree on a metric before we start.
  • If your data has no special rules and you need it now: a cloud service is quicker to start.

We would rather say no at the start than deliver a pilot that is of no use to you.

How it starts

Everything starts with the free feasibility study, which is the first phase of a closed-scope pilot, not an open-ended project.

  1. 1 / 3

    Free feasibility study

    You tell us the case and we review data, infrastructure and metric. We tell you whether there is a pilot before we start.

  2. 2 / 3

    Pilot on your server

    With your data and on your infrastructure, with a metric agreed before we start.

  3. 3 / 3

    Result and decision

    A working system, a report and the decision to continue, adjust or stop.

See the pilot in detail

Where we have it running

Dental Brain has been running on the server of a clinic with two sites since January 2026. Its vision model, trained with the NVIDIA Innovation Lab grant, is under clinical validation.

Ready-made platform

Looking for a ready-made platform to run your business?

Neural Fabrik

Frequently asked questions

01How much does it cost?

The feasibility study is free. The pilot depends on the scope, the data and the infrastructure: we give you a fixed price when the study ends, before we start.

02Does my data leave my organization during the pilot?

We work on your infrastructure. If any phase needed data to leave it, for example to train on external GPUs, we would propose it beforehand and in writing, with the data pseudonymised, as we describe in chapter 3 of the Sprint DGX series. Read the chapter

03What do I keep when the pilot ends?

The system stays running on your server and you can keep using it with our maintenance. The code remains ours.

04Who does the work?

We do, with no subcontracting: the team that listens to you is the team that builds.

05Which models do you use?

We use Gemma 4 as the base model, serve it with vLLM on an NVIDIA DGX Spark and adapt it to each case with fine-tuning. It is what we have tested and written about in the Sprint DGX series; in the feasibility study we see whether it fits your case.