Build · Private model and fine-tuning
A private model, only when the hosted model falls short
Evaluate first. Fine-tune an open model in your own cloud account only if the hosted model falls short.
- Who it is for
- A company whose hosted model may fall short on its task.
- Length
- Typically two to six weeks
What you get
A connection (endpoint)
The fine-tuned open model, available as a connection you can call.
A comparison report
Accuracy and speed, side by side with the hosted model.
Your cloud account
The model runs in your own cloud account.
Standalone or after a sprint
Buy it on its own, or after an AI production sprint.
Your environment
The model runs in your own cloud account. We do not train on your data without written consent.
How it works
Evaluate first
The hosted model is evaluated on your task before anything else.
Fine-tune only if it falls short
An open model is fine-tuned only when the evaluation shows the hosted model is not good enough.
Compare and hand over
You receive the connection (endpoint) and a report on accuracy and speed.
What you keep
A connection in your cloud account, and a comparison report that shows whether it is worth it.
Questions
- When do we need a private model?
- When the evaluation shows the hosted model falls short on your task.
- Where does it run?
- In your own cloud account.
- Can we buy it without a sprint?
- Yes. It can be bought on its own, or after an AI production sprint.
- What decides whether it is worth it?
- An evaluation set built from your cases. The comparison report shows whether the private model beats the hosted one on those cases.
Private models
Our fine-tuned model against a large general-purpose model
We fine-tune a small model on one sample task and measure it against a large general-purpose model that only gets instructions, with no fine-tuning. Both are scored on the same test set.
3.5×
better reply quality
~16×
better key fields
~50×
faster replies
~125x
estimated lower running cost per ticket (internal estimate)
Reply quality as the training set grows
We fine-tuned a small model on a sample business task in our internal testing, trained it on more and more examples, and scored each version on the same held-out test set. A large general-purpose model gets only a prompt.
+252%
higher reply quality
3.5×
times the large model's reply quality
Our fine-tuned model at the largest training set, against the large model.
- Our fine-tuned model
- A large general-purpose model
| Training examples | Our fine-tuned model | A large general-purpose model |
|---|---|---|
| 350 | 0.130 | 0.225 |
| 700 | 0.369 | 0.225 |
| 1,400 | 0.649 | 0.225 |
| 2,800 | 0.771 | 0.225 |
| 5,600 | 0.792 | 0.225 |
Results on a support-ticket task. Your results depend on your data.
Sample task: support tickets. Synthetic test set held out from training. The large model got instructions and no fine-tuning. The operations comparison is an internal estimate from compute time, not a measured invoice.
Next step
Bring one process.
30 minutes. An honest answer on whether AI fits.