Custom model training

Custom model training: LLM fine-tuning for one high-volume task

Lower running costs for one task you run at high volume, without giving up accuracy. We fine-tune an open-weight model on your labelled cases for that task, then test it against your current model on cases it has never seen, and you switch only if it costs less and is at least as accurate. The model can run on your own servers, so the data stays in the company.

Our team has built software and production AI for

trivagoSAPToniesEWEtecRacer360 Health & PerformanceConvointentFig Mayo

What a narrow task needs from a model

Frontier models are trained to handle almost any request. A task such as classification, extraction, scoring or routing needs one behaviour, done reliably. Cases you have already labelled teach a smaller model exactly that.

Your past cases

Records where the right answer is already known, such as confirmed fraud, corrected extractions or resolved tickets.

Labelled examplesReference test set

Base model

An open-weight model with a few billion parameters. It fits on a single GPU, and its licence allows commercial use.

Open-weight base modelLicence check

Fine-tuning

The base model learns from your known outcomes through supervised fine-tuning or reinforcement learning, with LoRA adapters.

LoRA adaptersReinforcement fine-tuning

Comparison

The model you use today and the new one answer the same cases, none of them seen in training. We compare accuracy, latency and cost per request.

Held-out casesCost per request
Goes live only if it is cheaper and at least as accurate

What a training project involves

On rented GPUs, a training run finishes within hours. Most of the effort goes into the training data and the test set, and into operating the model in your infrastructure once real requests reach it.

Data and tests

Labelled examplesReference test setBaseline scores on your dataMistakes reviewed with your experts

Model training

Open-weight base modelLoRA adaptersReinforcement fine-tuningLicence check

In production

Serving with vLLMFallback to a hosted modelDrift monitoringRetraining plan

When a task is worth its own model

The thresholds below come from our own cost models, and we check them against your real traffic.

Volume

A single task that receives several hundred thousand requests a month

Labels

The true outcome of each case becomes known later

Control

Strict response-time limits, or data that must stay on your own systems

Below these thresholds, a smaller hosted model is the cheaper choice, and that is what we recommend.

Talk to an engineer

How the two models compare on your cases

We run both models on one set of held-out cases and report accuracy, latency and cost next to each other. The trained model takes over only if it is cheaper to run and loses no accuracy.

Frontier model compared with a trained model

Example with synthetic data

Cost of 1000 requests
€4.80 to €0.70
Median latency
1.4 to 0.6 s
Recall
88% to 91%

At 2 million requests a month, the cost falls from €9600 to €1400

Your data and the model stay under your control

We take no payments from model or tool vendors. Our recommendations depend only on how each option performs on your task.

Your training data does not leave your environment.

You own the weights of the trained model.

We record test results and decisions so your auditors can follow them.

Where you have a works council, we deal with its requirements early on.

Questions

Do you charge for the first call?

No. You talk to one of our engineers for 30 minutes about the task you have in mind. You don't share any data, and we don't need access to your systems.

Is there evidence that a small trained model can beat a frontier model?

Yes, for narrow tasks with good labels. In October 2026, Coinbase reported that a Qwen3.5-9B model trained on its own fraud outcomes did better than a frontier model, with less than half the latency. For open-ended reasoning, or tasks that draw on broad knowledge, frontier models remain the better choice.

Coinbase engineering write-up

What if the trained model falls short of our current one?

Then you keep your current model. The test set and the evaluation pipeline stay with you, and you can use them to choose and monitor any model you run later.

Have you trained and run models in production before?

Yes. Our engineers trained computer vision models on more than 100 million images and ran them for a website with millions of visitors a day. We still work in the order we used then: first the test set, then the training, then monitoring once the model is live.

What do you need from us?

A task with enough volume and earlier cases whose outcomes you know, ideally several thousand of them. We also need one or two hours a week from a domain expert who can judge the model's answers.

What if we can't share production data?

Then training starts from open datasets or synthetic data. Or we begin with model selection: we compare existing models on your task and train only if it still pays off.

Which base models do you work with?

Open-weight models whose licences allow commercial use, such as Qwen, Gemma, Mistral, Granite and gpt-oss. The terms differ between models. Some exclude companies above a certain size or companies based in the EU, so we check them for every project. If your procurement rules require it, we use models of European or US origin.

How does the EU AI Act affect a fine-tuned model?

A LoRA fine-tune often uses far less compute than the original training, but the method alone does not decide your role under the AI Act. The Commission's guidelines name a modification using more than a third of the original training compute as an indicative criterion for when a company that modifies a model becomes the provider of a general-purpose AI model. An indicative criterion does not settle every case. Your role and obligations depend on how you actually modify the model and how you use it. We prepare the documentation your data protection and compliance teams need, including how the model was modified and the compute it used, and the legal assessment remains with your counsel.

European Commission

Talk to an engineer

Bring one piece of work you would like to improve. The call takes 30 minutes and needs no access to your systems or your data.

Booking runs on Google Calendar.

Mickey Graf

Mickey Graf

Founder and CEO of Cloudsail

Rated 4.9 out of 5 on Clutch

Prefer to write first? Leave your email and a few lines. Mickey reads every message himself and replies by email with times for a call.

The button prepares an email in your own email program, for you to send. Privacy policy (draft)