Model distillation

Pay less per call, with a model you own

We distill your expensive LLM calls into a small open-weight model, trained on your own traffic and tested against the model you use today.

€2,900 per task, fixed price. You pay only if it matches your current model on your own held-out examples.

Tell us about your calls

How it works

1
You send a sample of real calls
Inputs plus the outputs your current model gives. If you only have the inputs, we label them with your current model where its provider’s terms allow it, or with an open-weight model where they do not.
2
We train a small model and test it
We train a small open-weight model on your sample, then test it against your current model on examples it never saw.
3
You see the numbers, then decide
A written report puts quality and per-call cost side by side. You pay only if it matches.

Is your task a fit?

A good fit

A repetitive task you run at volume:

Classification
Extraction
Routing
Summarization
Structured output

Not a fit, and we will tell you so

If your task is one of these, we say so plainly:

Open-ended reasoning
Tasks that change every week
Low volume, where the savings would not cover the fee

Run it your way

Your choice, once the model matches.

The weights and a serving config

You get the model weights plus a ready serving config, and run it anywhere you like.

We host it

We run it for you, for a separate monthly fee based on your volume.

Questions

What do you need from us?: A sample of real calls: the inputs, and the outputs your current model gives for them. If you only have the inputs, we label them with your current model where its provider’s terms allow it, or with an open-weight model where they do not.
Which models are involved?: On your side, the model you call today: we work from its inputs and outputs. On ours, a small open-weight model trained on your sample. You take the weights, or we host it for you.
Can you train on my current provider’s outputs?: Only where your provider’s terms allow it. Some providers restrict training other models on their outputs, so we check your provider’s terms with you before we start. Where they restrict it, we label your sample with an open-weight model instead, and the comparison with your current model still decides whether you pay.
How is “matches” decided?: Before any work starts, we agree with you in writing on what counts as a match for your task. The small model is then tested against your current model on your own held-out examples, which it never saw in training. If it does not meet that criterion, you do not pay.
What happens to our data?: It is used only to build your model, and deleted when the project ends.

Price

€2,900 per task, fixed price

You pay only if the small model matches your current model on your own held-out examples.

What counts as a match is agreed in writing before work starts.

Your data is used only for your model and deleted when the project ends.

Start here

Tell us about the calls you want to make cheaper

A few lines is enough. You get a personal answer within one business day, from the engineer who would do the work.

How it works

You send a sample of real calls

Inputs plus the outputs your current model gives. If you only have the inputs, we label them with your current model where its provider’s terms allow it, or with an open-weight model where they do not.

We train a small model and test it

We train a small open-weight model on your sample, then test it against your current model on examples it never saw.

You see the numbers, then decide

A written report puts quality and per-call cost side by side. You pay only if it matches.

Questions

What do you need from us?

A sample of real calls: the inputs, and the outputs your current model gives for them. If you only have the inputs, we label them with your current model where its provider’s terms allow it, or with an open-weight model where they do not.

Which models are involved?

On your side, the model you call today: we work from its inputs and outputs. On ours, a small open-weight model trained on your sample. You take the weights, or we host it for you.

Can you train on my current provider’s outputs?

Only where your provider’s terms allow it. Some providers restrict training other models on their outputs, so we check your provider’s terms with you before we start. Where they restrict it, we label your sample with an open-weight model instead, and the comparison with your current model still decides whether you pay.

How is “matches” decided?

Before any work starts, we agree with you in writing on what counts as a match for your task. The small model is then tested against your current model on your own held-out examples, which it never saw in training. If it does not meet that criterion, you do not pay.

What happens to our data?

It is used only to build your model, and deleted when the project ends.

Tell us about the calls you want to make cheaper

A few lines is enough. You get a personal answer within one business day, from the engineer who would do the work.