Two sizes and individual projects
Offering
Price on request. For M and L your quote is in your customer account right after the request; for individual projects we prepare it from your description.
| M | L | Individual | |
|---|---|---|---|
| Price | on request | on request | on request |
| Base model | Qwen3.5-9BApache 2.0 · Open Source | Qwen3.8-27BApache 2.0 · Open Source | as required, other base models possible |
| Data volume | ZIP upload up to 8 GB | ZIP upload up to 16 GB | by agreement |
| Model size | 9B1 | 27B1 | larger or tailored differently |
| Suitable for | Balanced model for most departments and medium data volumes. | Most capable model for demanding tasks and large knowledge bases. | Larger models, several models, special requirements for data, operation or integration.Send a request |
M
- Price
- on request
- Base model
- Qwen3.5-9BApache 2.0 · Open Source
- Data volume
- ZIP upload up to 8 GB
- Model size
- 9B1
- Suitable for
- Balanced model for most departments and medium data volumes.
L
- Price
- on request
- Base model
- Qwen3.8-27BApache 2.0 · Open Source
- Data volume
- ZIP upload up to 16 GB
- Model size
- 27B1
- Suitable for
- Most capable model for demanding tasks and large knowledge bases.
Individual
- Price
- on request
- Base model
- as required, other base models possible
- Data volume
- by agreement
- Model size
- larger or tailored differently
- Suitable for
- Larger models, several models, special requirements for data, operation or integration.Send a request
What model size determines
The volume of your documents is the limit, not the reason. What matters is what the model is meant to do with what it learned.
M · 9 billion parameters
One clearly defined area with consistent terminology: one product line, one department, one set of rules. Answers that follow directly from what was learned.
L · 27 billion parameters
Several areas in one model, inconsistent or contradictory sources, several languages. Tasks across multiple steps where the model has to connect things rather than look one up.
This capability comes with model size and cannot be replaced by more training data. If in doubt, decide by the task rather than by the volume.
You accept the quote in your customer account. A quote is valid for 30 days. When processing starts you receive one invoice for the quoted price, net, payable within 14 days.
You source and choose the hardware yourself; as availability and market prices fluctuate, this keeps you free to buy at the best available price.
¹ Model size in billions of parameters.
What the price covers
What we do on the way to your AI
The price covers more than compute. Until your model is ready, we take on four stages and stand behind each of them.
Preparation
We check the archive, extract the text from Office files, PDF, emails, HTML and images, and remove personal data with a model on our own instance. Then we condense the material so that duplicates and irrelevant content do not end up in the model.
Preparation
We check the archive, extract the text from Office files, PDF, emails, HTML and images, and remove personal data with a model on our own instance. Then we condense the material so that duplicates and irrelevant content do not end up in the model.
Dataset
From the material we generate training examples of a fixed size: documented facts from your documents and tasks that preserve general language ability. A sample is checked; if it does not pass, we generate the dataset again. No external AI service sees your documents.
Dataset
From the material we generate training examples of a fixed size: documented facts from your documents and tasks that preserve general language ability. A sample is checked; if it does not pass, we generate the dataset again. No external AI service sees your documents.
Fine-tuning
For each order we set up a dedicated GPU instance. There the training adapter learns your knowledge while the base model remains unchanged. We then prepare the model for operation on a single graphics card and check the result. Afterwards we destroy the instance and its storage.
Fine-tuning
For each order we set up a dedicated GPU instance. There the training adapter learns your knowledge while the base model remains unchanged. We then prepare the model for operation on a single graphics card and check the result. Afterwards we destroy the instance and its storage.
Delivery
You receive the model weights, the training adapter, an ownership certificate with SHA-256 checksum, the licence texts and a verification tool for the labelling, plus the setup program for Windows, Linux and macOS with a browser interface. The package is stored encrypted for 30 days and unlocked with your activation code.
Delivery
You receive the model weights, the training adapter, an ownership certificate with SHA-256 checksum, the licence texts and a verification tool for the labelling, plus the setup program for Windows, Linux and macOS with a browser interface. The package is stored encrypted for 30 days and unlocked with your activation code.
What we take on for you
We take on these obligations with every order. They are part of the price.
Processing under Article 28 GDPR.
We conclude the agreement with you, document our safeguards, control our sub-processors, grant you audit rights and notify you of a data breach within 48 hours.
Obligations under the AI Act.
We label every output with signed, machine-readable metadata under Article 50(2) and supply the tool that verifies the signature independently of us.
Licence and ownership.
We meet the licence obligations of the base model (Apache 2.0) and, on payment, transfer to you all transferable rights in the result, evidenced by the certificate.
Processing in the EU.
We store in Germany and compute at a provider named in the data processing agreement in the EU or Switzerland. No external AI provider gains access to your documents.
Security of processing.
We take your upload straight into encrypted object storage, encrypt every artefact with its own key and destroy the instance after the order.
Warranty.
We stand behind the described procedure and the setup program running on compatible hardware under statutory warranty. If data is unusable, we give you the opportunity to supply suitable data.
Included in the price
What is included in every tier
The following points apply equally to both tiers. The full terms are set out in our terms and conditions.
One payment, then your property.
One advance invoice for the price in your quote, net and payable within 14 days. No recurring licence or usage fees follow.
No limit on users.
The model runs entirely on your own hardware, locally. How many staff work with it is your decision, at no extra cost.
Full rights to the model.
You may use, modify, reproduce and commercially exploit it without limitation in time, territory or scope, including reselling it. You only need to observe the Apache 2.0 terms of the base model.
Certificate of ownership with checksum.
You receive a certificate with cryptographic proof (SHA-256) for the model weights.
Deletion of our copies.
We delete our copies of the model and the training data no later than 30 days after delivery. Once deletion has taken place, you automatically receive proof of deletion.
Balance-sheet treatment as an asset.
The model is a purchased intangible asset and, in principle, capable of being capitalised and depreciated; what you recognise are the acquisition costs, that is the price you paid. Please align the specific accounting treatment with your tax advisor.
Fine-tuning and document retrieval (RAG) are not alternatives, they solve different jobs. Differences and limits in the knowledge section.