A custom model fine‑tuned on your examples, for one task at volume
Fine-tuning and distillation on open-weight models, for the task a general model does too slowly, too expensively or not quite right. Scored against the general model before it ships.
No list price. It starts with a diagnostic at a fixed fee, and the build is quoted in writing after it.
- Invoice, scannedPayables
- Receipt, photoExpenses
- Statement, PDFBank
When you need a custom model
And when you do not.
Fine-tuning and distillation on open-weight models, for one bounded task where a general model is too slow, too costly or not quite right. It is the exception, not the first step, and the first section says when it is the wrong tool.
You need it when
- One task at high volume, where a general model costs too much for each call.
- A format or house style a general model does not hold reliably, with a few hundred to a few thousand checked examples to learn from.
- Latency a hosted model cannot meet, or a model that has to run on your own hardware.
- A large model that works, and a smaller one you want to carry the same task at a fraction of the cost.
You do not when
- New knowledge. A model does not learn your documents by training on them; retrieval does that, and keeps the citations.
- A task prompting already does well. We measure first, and often the answer is a better prompt.
- Fewer than a few hundred good examples.
What a custom model includes
Part by part.
- 01
The task is bounded
One input, one output, written down with examples of right and wrong.
- 02
The baseline is measured
Prompting and retrieval on a general model, scored on a held-out set. This is the number to beat.
- 03
The training set is built
Your examples, cleaned, checked and split, with a held-out set nobody trains on.
- 04
A model is adapted
An open-weight model fine-tuned with a low-rank adapter, or a smaller model distilled from a larger one.
- 05
It is scored against the baseline
Same held-out set, same measures. It ships only if it wins by the margin agreed.
- 06
It is served
Through private inference in your account or on your hardware, with the baseline kept as a fallback.
How a custom model is proved
A locked test set and a pass mark, agreed before the build.
- Accuracy on a held-out set the model never saw, against the general-model baseline, with the margin agreed before training.
- Cost and latency for each call, measured in the place it will run.
- Drift: the same set re-scored on a schedule, with an alert when the number moves.
How a custom model is run
After launch: monitoring, updates and what you own.
- Monitoring on every call: errors, latency and the share routed to the fallback raise an alert.
- Retraining when the data or the task changes, scored on the held-out set before release.
- Handover: the weights, the training set, the scoring code and a runbook, in your name.
What a custom model will not do
Limits we set on purpose.
- It does not train on your data for anyone else, and it does not train for you without a written agreement.
- It does not replace retrieval for facts that change.
- It does not ship if it does not beat the baseline.
- It is not a general assistant. One task, done well.
Questions about a custom model
Will our data train a model for anyone else?
No. A model trained on your examples is yours: the weights, the training set and the scoring code are handed over in your name. Nothing you give us trains anything for anyone else.
Why try prompting and retrieval first?
Because they usually get there, and they are cheaper to change. Training a model is worth it for one bounded task at volume, a format a general model cannot hold, or latency and hardware limits. We measure the baseline first and show you the number.
How many examples do we need?
A few hundred checked examples for a narrow task; a few thousand for a wider one. Fewer than that, and we would rather improve the prompt.
Which models can be trained?
Open-weight models whose licence allows your use, chosen for the task and for where they will run. We say which, and why, in the quote.
Talk to us about custom models
Fifteen minutes on a call is enough to tell whether it fits your work. If it does not, we say so.