Custom Model Training
More accurate than frontier models, at a fraction of the cost
We retrain open-source models such as Qwen, Gemma and Mistral on your data, one task at a time: classification, extraction, scoring or routing. On tasks like these, a trained model of a few billion parameters regularly outperforms frontier models on accuracy, at a fraction of the inference cost, and it runs in your infrastructure. Every model is measured against your current one on your own test set before it ships.
Our team has built software and production AI for
Why task-specific training wins on narrow tasks
Frontier models are trained for breadth. A narrow task with known historical outcomes needs one behaviour done reliably, and your labelled data is the training signal for exactly that.
Your data
Historical cases with known outcomes, such as confirmed fraud, corrected extractions or resolved tickets.
Base model
An open-source model of a few billion parameters, licensed for commercial use and sized for a single GPU.
Post-training
Supervised fine-tuning or reinforcement learning against your known outcomes, usually with LoRA adapters.
Evaluation
Compared with your current model on the same held-out cases for accuracy, latency and cost per request.
Everything around the model
Training itself takes hours on rented GPUs. Most of the work is in the training data, the test set and operating the model once it serves traffic.
Data and testing
Training the model
Running it
When custom training pays off
Indicative thresholds from our own cost models. We validate them against your traffic.
Volume
Several hundred thousand requests per month on a single task
Labels
Ground truth becomes available after the fact
Control
Latency requirements, or data that has to stay in your infrastructure
Below these thresholds a smaller hosted model is usually the more economical choice, and we recommend it.
Talk to an engineerMeasured on your data, not on benchmarks
Your current model and the trained model run on the same held-out cases. Accuracy, latency and cost are reported side by side, and the trained model ships only if it wins.
Frontier model against your own model
Sample result, synthetic data
€9600 to €1400 a month at 2 million requests
Your data stays yours
We are independent of model and tool vendors. No vendor pays us, so our recommendations follow what works for you.
Training data stays in your environment.
The model weights belong to you.
Test results and decisions documented for your auditors.
Works council requirements handled early.
Questions
What does the first call cost?
Nothing. It is a 30-minute conversation with one of our engineers, with no access to your systems and no data shared.
What evidence is there that trained small models outperform frontier models?
On narrow tasks with good labels, the published results point the same way. In October 2026 Coinbase reported that a Qwen3.5-9B model trained on its fraud outcomes outperformed a frontier model at less than half the latency. Open-ended reasoning and tasks that need broad knowledge stay with frontier models.
What happens if the trained model does not outperform the current one?
It does not ship. You keep the test set and the evaluation pipeline, which remain useful for selecting and monitoring any model you use.
Have you trained models in production before?
Yes. Our engineers trained and operated computer vision models on more than 100 million images for a website with millions of daily visitors. The method carries over: test set first, then training, then monitoring in production.
What do we need to provide?
A task with sufficient volume, historical cases with known outcomes (ideally several thousand), and one or two hours a week from a domain expert who can judge the outputs.
Which models do you use?
Open-weight models licensed for commercial use, including Qwen, Gemma, Mistral, Granite and gpt-oss. Licence terms vary, and some exclude companies above a size threshold or based in the EU, so we review them per project. Where procurement requires it, we use models of European or US origin.
What about the EU AI Act?
Fine-tuning with LoRA uses a small fraction of the original training compute, so under the Commission's guidelines you would not normally become the provider of a general-purpose AI model. Obligations still depend on the use case. We prepare the documentation for your data protection and compliance functions, and the legal assessment stays with your counsel.
Talk to an engineer
A 30-minute call with one of our engineers about the task you have in mind, what it costs today and whether a custom-trained model could take it over. No access to your systems and no data shared.
Check your email program
We tried to open a draft in your email program. Nothing is sent until you send it, and this page cannot tell whether the draft opened or the email was delivered. The form stays editable, and you can also write to miki@cloudsail.com directly.