Skip to content
Measured in watt-hours

Computing power

The right size

Why, for most tasks, we use the smallest model that solves them reliably. And how we measure that instead of assuming it.

Question 1

Does the task need a language model at all?

Sorting texts, for example. A classical method that has been around for decades can do that too.

0.8 secondsversus3 hours 48 minutesThat is how long a 9-watt LED lamp would have to burn to use the same electricity. For the same work, the same batch of texts.

Classical method0.0021Wh

On this scale that is two pixels, and even those are drawn too large.

Language model, 70 billion parameters34.15Wh

16,000times more electricity for 0.05 points more accuracy.

Measured for one run over the same test dataset. Source: Comparing energy consumption and accuracy in text classification inference, Scientific Reports, 2026. The lamp time is calculated from it, at 9 watts.

Question 2

If so, how large does it need to be?

The same model, small and large, the same task.

7 billion parametersone seventh
72 billion parametersseven sevenths

7times more electricity for 0.07 points more accuracy.

We use the smallest model that solves the task reliably.

Reliably means: it makes no mistakes that go unnoticed. We check that against your own documents before anything goes live.

When we do use a large one

So that this doesn’t sound like penny-pinching: there are tasks where a small model is the wrong choice.

  • Free text that looks different every time.
  • Images, drawings and photos, especially when they carry dimensions.
  • Several languages in one document.
  • Wherever a small model would make mistakes that nobody notices.

Then we work with the large providers, and we tell you which one.

What this means for you

Your data

In your systems

The assistant works in the files and programs you use today. Everything else runs in a separate instance of its own on servers in the EU, or on a machine on your premises if you prefer. No training with your data.

The price

Less to run, lower cost

A small model costs a fraction to run. That does not sit in our margin, it sits in your quote.

The consumption

Frugal from the start

Less computing power means less electricity, every day and with every request. That is decided when we build it, not offset afterwards.

Our contribution

The most frugal electricity is the one never used.

Every assistant we build runs on the smallest model that solves its task reliably. That cuts consumption where it arises: in the running of your assistant, every day, with every request. Not afterwards, not somewhere else.

In the quote
Which model we use and why it is enough
In operation
A separate instance of its own on servers in the EU
On request
Estimated consumption per project, to pass on

When your client asks

As a mid-sized company you are not required to report: the threshold is a thousand employees. Your large clients ask anyway.

For every project we provide, on request, one page with the model size, the data centre location and the estimated consumption.

No expert opinion, no advice on reporting duties. Just the figures for the part we run for you, so that you can pass them on.