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
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.
Base model
An open-weight model with a few billion parameters. It fits on a single GPU, and its licence allows commercial use.
Fine-tuning
The base model learns from your known outcomes through supervised fine-tuning or reinforcement learning, with LoRA adapters.
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.
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
Model training
In production
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 engineerHow 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
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.
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.
When to retrieve and when to fine-tune
Retrieval supplies facts that change, such as current prices, policies or product documentation. The model looks them up when it answers, so an update needs no new training run.
Fine-tuning teaches repeatable behaviour. A model that sorts incoming claims into your categories, or pulls the same fields from every invoice, learns that pattern from your past cases.
Related services
AI agent evals
Find out how often an agent gets the job done before it acts in your systems, by checking outcomes, tool calls and how reliably runs repeat.
AI automation
Routine work taken off your teams: processes automated step by step with rules, workflow tools, models and agents, approved by people and measured per case.
AI chatbots
Customers and staff answered faster from your data, with approved actions and a handover to people when it matters.
AI strategy and rollout
AI strategy and rollout that starts with one running, measured system on your own processes and builds the roadmap around its figures.
Data integration for AI
Answers from file shares, SharePoint, ERP and tickets in seconds, for assistants and agents, with permissions carried through and tested.
AI engineers for your team
Extra AI engineering capacity when your team lacks AI skills: experienced engineers who build with you and leave the know-how.
AI evals engineering
Evals built from your own traces show whether each change made your AI product better or worse, with a release gate that stays in your repository.
AI system testing
Before go-live, acceptance tests on your own cases tell you how often your chatbot, assistant or agent is right, in terms your QA team already uses.
Free AI usage review
One page for management on where your AI money goes, whether it pays off and what to change, from exports you choose to share. Free, no system access.
Enterprise AI
One approved AI platform for the whole company: an internal assistant instead of shadow AI, approved models, data that stays in-house and spend per team.
LiteLLM gateway
Budgets and access rules for every team's AI use, through a LiteLLM gateway that is hardened, connected to your identity provider and kept up to date.
AI platform for all teams
One approved route to AI for every team, with budgets, logs and a free choice of model, run on your servers or in your cloud.
System access for agents (MCP)
AI agents that look up and update data in your systems, with narrow permissions, human approval for changes and tests.
On-premise AI
AI on offers, contracts and customer files, with the data staying in your company: own servers or own cloud account, tested before go-live.
Azure OpenAI and Bedrock
GPT and Claude on company data, under the cloud contracts you already have: Azure OpenAI and Amazon Bedrock in your own account, with EU processing and one gateway.
Knowledge assistants (RAG)
Staff find answers in internal documents in seconds, with sources cited, permissions kept and tests on your own questions.
RAG evaluation
An assistant that answers from your documents, tested on retrieval and on answers in each language your users write in, so you know how often it is right and cites the right source.
vLLM on Kubernetes
Open-weight models on your own GPUs, so data stays in house and costs follow capacity you plan, sized from measured load.
From the blog
RAG or fine-tuning?
Decision criteria for choosing retrieval, fine-tuning or both: whether the task needs changing knowledge or consistent behaviour, with a worked table and an order of work.
Fine-tuning data quality
Four checks that show whether your examples are ready for fine-tuning: label errors, coverage of real traffic, provenance and leakage into the test set.
Fine-tuned model evaluation
How to show that a fine-tuned model should replace the current one: acceptance criteria written in advance, a fair held-out comparison, and cost and latency in the same table.
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
Founder and CEO of Cloudsail
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.
Look for the draft in your email program
Send it from there, and Mickey replies in person with times for a 30-minute call. If no draft has opened, write to us at miki@cloudsail.com.