Skip to main content

Own models

A model that has learned how you work.

A good prompt explains your way of working to a stranger, every single time. Your own model has learned it. We train an open base model further on your knowledge — the dataset is yours and outlives every model swap.

The difference

Explaining, looking up, or knowing.

There are three places a rule can live. Only the third one survives a model swap.

In the prompt

holds until the next window

In the weights

holds without anyone writing it down

In memory

holds when someone searches for it

The dataset is the asset, not the model

A trained model ages, like every model. The dataset it came from does not. When a better base model arrives next year, we train again from the same dataset. That is why it is yours, in writing.

Your data never enters a shared model

Every customer gets their own training run and their own weights. There is no StudioMeyer model assembled from everybody's customer data. That is not a setting, that is the architecture.

No vendor lock-in

The base model is openly licensed, the result runs on your hardware or in your cloud. You do not need us to operate it. You need us if you want to develop it further.

Where it fits

Most of the time you do not need training.

Finetuning is for form, not for facts. There are four rungs, and each one only comes into play once the one below it has measurably failed, not felt wrong.

  1. 1

    Prompt

    The rule sits in the instruction. Costs nothing, changes instantly, holds until the next context window.

  2. 2

    Memory

    Knowledge that changes belongs here: prices, stock, customer history. Look it up rather than learn it in.

  3. 3

    Training

    Only once form, domain vocabulary or tool calls go wrong reproducibly. Then behaviour goes into the weights.

  4. 4

    Distillation

    A large model teaches a small one how to answer. For when it has to run small, fast and cheap.

We check first whether a good prompt and a memory are enough. In most cases they are, and then we say so, instead of selling a training run.

Four ways in

Where a rule lands decides whether it holds.

Not every rule belongs in the weights. Some are right in the prompt, some in memory, some need a hard guard outside the model. We sort that out before the first training run.

Knowledge

What your business knows: products, pricing logic, edge cases, the names that only mean something in your shop. Today that sits in a manual beside the model and gets sent along with every request.

After training the model knows it. Answers get shorter, faster, and less dependent on whether the right passage was retrieved.

Processes

How work is done at your place: in which order, with which intermediate steps, when a human has to sign off. Written out as a rulebook this gets longer than it helps.

As a learned sequence it needs no list. The model does it the way it was done in your examples.

Behaviour

How it answers: tone, length, limits. What must never be promised. When to ask instead of guessing. Exactly the part of a manual that gets ignored first.

Training behaviour means preferring and rejecting examples, not writing rules. Whatever has to hold absolutely, we additionally guard outside the model.

Robot connection

From the system's point of view a robot is just another client calling tools — like a chat window, only with arms. The interface is the same one we already build for connectors and systems.

The robot leaves eventually, the learned knowledge about your business stays. That is why we build it next to the robot, not inside it.

Straight

What training solves, and what it does not.

It solves this

  • An exact output schema

    The same form, the same JSON structure, ten thousand times, without outliers.

  • Narrow domain vocabulary

    Industry language, internal shorthand, standards, without anyone re-explaining them on every call.

  • Brand voice

    Consistent, without a three-page prompt in front of every answer.

  • Reliable tool calls on small models

    In practice the most important one: small models only get tool calls right reliably after training.

It does not solve this

  • Fresh or changing knowledge

    Prices, stock, appointments belong in memory, not in the weights. What changes gets looked up.

  • Small does not become clever

    A small model becomes narrower and more reliable, not smarter. Anyone claiming training turns an 8B model into a frontier model is selling you something false.

  • Learning has side effects

    A model forgets elsewhere when it learns something new. You work against that with data mixing, but you have to know about it.

  • Without a held-out test set nobody knows anything

    Without set-aside examples there is no way to tell whether a run improved something or broke it. This is where most projects fail.

With what

What we build it with.

Open models, traceable methods, nothing secret. Anyone who knows the stack can check every number.

Base model
An openly licensed model that permits commercial derivative work (Apache 2.0). The choice follows the task and the hardware, not fashion.
QLoRA + SFT
Our standard case, covering roughly 95 percent of projects: output format, domain vocabulary, voice and tool calls. We train an adapter, not the whole model. That keeps runs affordable and allows several specialisations on one base.
DPO / SimPO
Stage two, once supervised training holds. Sharpened on pairs following the pattern "A is better than B". Behaviour comes from this stage, not the first.
GRPO / DAPO
Reinforcement learning with a verifiable reward, in our case for the orchestrator. Only worthwhile when correctness is machine-measurable. With tool calls it is exactly that.
Own hardware or rented GPU
The run happens where the data is allowed to be. Locally for sensitive datasets, otherwise on rented GPUs that get wiped afterwards.
Evaluation
Checked against a test set from your actual work before delivery, not against a paper benchmark. No result without a comparison against the untrained model.

How it runs

Eight steps.

  1. 01

    Define the target behaviour, build the test set first

    Before the first line of training data comes the question: how will we know it worked? A set of examples is put aside and never touched during training.

  2. 02

    Check whether it works without training

    Prompt and memory are measured against that same test set. If that is enough, the project ends here, in your favour.

  3. 03

    Build the dataset

    The most honest point on this page: in company projects, 30 to 50 percent of total cost is pure data preparation. Examples come from real operations, carry origin and date, and are versioned like source code.

  4. 04

    Choose the method and the base model

    For roughly 95 percent of cases, QLoRA with supervised fine-tuning. The model itself stays untouched; learning happens in small additional matrices. The result is a file of a few hundred megabytes, not a new model.

  5. 05

    Train on rented compute

    No hardware of your own required. The card is rented by the hour and shut down afterwards.

  6. 06

    Measure against the held-out test set

    Before and after, with numbers. No gut feeling, no "it feels better".

  7. 07

    Ship

    One base model carries several specialists at once, each as its own small file. Instead of five models at 20 GB each, one model runs with five attachments.

  8. 08

    Maintain

    When a new model generation arrives, a new attachment is trained from the same dataset. That is exactly why the dataset is the asset and not the model.

FAQ

Frequently asked

Why not just write a better prompt?

Because a prompt applies anew on every request and ends with the context window. For many tasks that is enough, and then that is exactly what we recommend. Your own model pays off when the same explanation is sent a thousand times, or when a rule has to hold even when nobody writes it down.

Is it really your own model if you train an existing one further?

It is your own model in the sense that the weights are yours and came from your data. It is not a model built from scratch — nobody does that on a sane budget, and anyone claiming otherwise is selling you something else. The base model's licence explicitly permits commercial derivative work.

Does our data end up in a model other people use?

No. Every customer gets their own run and their own weights. There is no shared model made from customer data — not as a promise, but because that is not how we build it.

What does it cost?

The price follows the size of the dataset, not the compute time. Compute is the smaller line item. What costs time is sorting and preparing your material, and we can only see that once we have looked at it.

How do we know this works?

We run the method on our own orchestrator, the system behind the AI station. What we build for you runs in our own operation first. And it is measured against your own test set, not a paper benchmark: without a comparison to the untrained model we do not ship.

And what about n8n and workflow automation?

We build it wherever a process needs it. It is simply not on the shelf as a separate product; it comes as part of the system we build for you. The reason is simple: a single mail watcher does little if it does not know your business. That is what this page is about.

Next step

Which explanation do you send along every day?

Tell us what your model has to learn again every time. We will tell you honestly whether that belongs in training or in a better prompt.