Specialist models

Beat frontier models at a fraction of the cost

We train small specialist models on your own data. On the narrow tasks your business runs all day, such as fraud checks or document extraction, they can be more accurate than the largest frontier models and cost a fraction to run. If ours doesn't beat your current model on your own test set, it doesn't ship.

Talk to an engineer How it works

Our team has built software and production AI for

How a small model beats a large one

A frontier model knows a little about everything. Your task needs one thing done very well, thousands of times a day, and you already hold the answers it should learn from.

Your outcomes

Years of confirmed answers, such as fraud you caught or documents your team corrected.

Labelled historyGolden test set

A small open model

An open-weight model of a few billion parameters, licensed for commercial use and small enough for one GPU.

Open-weight base modelLicence check

Trained on your task

Fine-tuning, often with reinforcement learning against your known outcomes, teaches it the one job it has to do.

LoRAReinforcement fine-tuning

Tested against the frontier

It runs next to your current model on the same cases, and ships only if it wins on accuracy and cost.

Serving with vLLMRetraining plan

Or it doesn't ship

Everything around the model

Training takes hours on rented GPUs. Whether your model wins is decided by the test set and the training data, so that is where we spend most of our time.

Data and testing

Labelled historyGolden test setBaselines on your casesError review with your experts

Training the model

Open-weight base modelLoRAReinforcement fine-tuningLicence check

Running it

Serving with vLLMHosted fallbackRetraining plan

Where small models win

Rough rules from our own sums, not guarantees. We check them against your traffic.

Volume

Hundreds of thousands of requests a month on one task

Labels

You find out later what the right answer was

Control

Latency matters, or the data can't leave your infrastructure

Below that, a cheaper hosted model is usually the better deal, and we'll say so.

Talk to an engineer

Measured on your data, not on benchmarks

Your current model and the trained one run on the same cases. You see accuracy, latency and cost side by side, and the small model ships only if it wins.

Frontier model against your own model

Sample result, synthetic data

Cost per 1000 requests

€4.80 to €0.70

Median latency

1.4 s to 0.6 s

Recall

88% to 91%

€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.

Can a small model really beat a frontier model?

On a narrow task with good labels, yes. In October 2026 Coinbase reported that a fine-tuned Qwen3.5-9B beat a frontier model on its fraud task, with less than half the latency. It does not hold for open-ended work, which is why we test on your cases before anything ships.

Coinbase engineering write-up

What if it doesn't beat our current model?

Then it doesn't ship, and you have spent a few weeks rather than a budget. You keep the test set, which is useful on its own for choosing and checking any model you use.

Have you trained models in production before?

Yes. Our engineers trained and ran computer vision models on more than 100 million images for a website with millions of daily visitors, years before language models. The method has not changed: a test set first, then training, then watching the model once it runs.

What do we need to bring?

A task that runs often, and past cases where you know the right answer, ideally a few thousand. We also need an hour or two a week from someone who can tell a good answer from a bad one.

Which models do you use?

Open-weight models whose licences allow commercial use, such as Qwen, Gemma, Mistral, Granite and gpt-oss. Some licences exclude companies above a certain size or based in the EU, so we read them before we start. If your buyers would rather not use a Chinese model, we pick a European or US one.

What about the EU AI Act?

Tuning a model with LoRA uses a tiny fraction of the compute that went into the original, so under the Commission's guidelines you usually don't become the provider of a general-purpose model. What you use it for still decides your obligations. We prepare the paperwork for your data protection and compliance people, and your lawyers make the call.

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 specialist model could take it over. No access to your systems and no data shared.