Small private models
Small private models for narrow, repeatable jobs
Not every task needs a large language model. For extraction, classification, reranking and masking personal data, a small model built for one job runs on your own server, costs only the server, and keeps the data where it is.
What are small models good at?
Narrow jobs that repeat thousands of times, where the text is too varied for rules but the task never changes:
- Pulling specific data out of text that shouldn’t go to a cloud model
- Classifying tickets, documents or requests, and scoring them by category
- Reranking search results so the best passage comes first
- Masking names, ID numbers and account numbers before anything reaches a cloud model, then restoring them in the answer
Too messy for regex. Too simple for an LLM.
In our tests, a small model processes a prepared batch in under a second once loaded, and sorting and reranking run quickly even in a modest container. There are no per-token fees: the running cost is the server.
Can you train a model for our specific needs?
Yes, when no public model handles your documents or categories well enough. These are small, narrowly specialised models, not LLMs: encoders such as RoBERTa and BERT, named-entity models and the like. First we look on Hugging Face for a ready open model that already does the job. If nothing matches closely enough, we take the nearest one and fine-tune it on your examples — and when even that falls short, we train a model from scratch.
The trained model stays yours and runs where your data lives. A custom-model project is planned for up to 3 months.
Where we are
Small models are the main direction of our core. The results above come from our tests of public models, not yet from a client deployment where a small model replaced an LLM. Our first projects of this kind are the founding-client projects, published as case studies with real numbers.
| Rules / regex | Small private model | Large language model (API) | |
|---|---|---|---|
| Handles varied, unpredictable text | No | Yes | Yes |
| Speed per item | Instant | Under a second per prepared batch* | Seconds |
| Running cost | About zero | The server only | Per token, grows with volume |
| Data leaves your infrastructure | No | No | Yes, unless self-hosted |
| Best for | Fixed formats | Narrow, repeatable tasks | Reasoning, open questions, conversation |
*Measured in our tests on a prepared text chunk with the model already loaded; real throughput depends on the task and the hardware.
Questions about this service
Is our data used to train anyone else’s model?
No. Your data trains only your model, if you order one. Never ours, never a model vendor’s.
How long does a custom model take?
Up to 3 months, depending on the task and how many labelled examples exist. We start by testing whether a small model handles your documents at all.
What does it cost?
Custom-model projects are scoped and quoted individually after the first call, with a fixed price agreed before the work starts.
Bring us one workflow. We’ll tell you what it needs.
A 30-minute call. Bring one real process or one sample document. We’ll tell you whether it needs an LLM, a small model, or just a script, and what it would take. No deck.