The AI that belongs to you.

Yes. Upon full payment, all transferable rights to the fine-tuned model, including the model weights, pass to you irrevocably. We retain no rights of use, exploitation or recall and delete our copies after delivery. As proof you receive a certificate of ownership with a hash (SHA-256).

Yes. 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, essentially the licence notice; the licence text is included with the delivery.

The model becomes your property upon full payment and is in principle capable of being capitalised and depreciated as an intangible asset. What you recognise are the acquisition costs, that is the price you paid; the base model is licensed under Apache-2.0 and causes no separate cost. The certificate of ownership and the invoices serve as evidence. Please align the specific accounting treatment with your tax advisor.

Nothing changes for operation. Your model runs entirely locally on your own hardware, with no licence server or cloud account. Everything you need to operate it is in your building and belongs to you.

Your documents are used exclusively to train your model and are stored in the EU or Switzerland. The processing takes place on a GPU instance provisioned solely for your order in the EU or Switzerland, run by our own software, without any external AI provider; we obtain your consent before starting.

Yes. As soon as you confirm via the link we send you that your model has been made available for download, we permanently delete uploaded documents, the training data derived from them and our copies of the model. Without confirmation we remind you twice; 30 days after provisioning we delete even without confirmation. We retain order and invoice data only within statutory retention periods.

Among others PDF, Word, Excel and PowerPoint, as well as text, Markdown and HTML files, bundled as a single ZIP file. Unusable files are skipped and logged; the order continues.

There is no deposit. Once processing of your data begins, we issue one invoice for the full model price; it is net and payable within 14 days.

Once your order is accepted and your data is uploaded, the pipeline starts automatically. For M, preparation, training and quality checks usually take just a few days, for L we expect around a week. Once your model is ready, you receive the activation code by email and can download it.

Yes, you source the hardware yourself; we do not supply a device. Availability and prices on the hardware market currently fluctuate considerably. Rather than tie you to a specific device or price, we leave the choice to you: you buy at the best available price and are not bound to our procurement. We recommend compatible devices with matching specifications, for example the NVIDIA DGX Spark or the DGX Station; you load the model onto it using our program.

On open Qwen models under the Apache 2.0 licence: Qwen3.5-9B for size M, Qwen3.8-27B for size L. That is exactly why we can transfer all rights to the fine-tuned model to you; this is documented in the certificate of ownership, which names the base model.

Each solves a different task. Document retrieval (RAG) searches stored documents and keeps individual facts current, but knows neither your use case nor your company's way of working or language. Your fine-tuned model has learned exactly that. The right combination uses both: stable long-term knowledge in the model, maintainable factual knowledge in retrieval. AnythingLLM, the interface provided with your model, already includes document retrieval; you switch it on at no extra cost and manage your documents yourself.

No special IT expertise is required. You download the setup program from our website, install it on your hardware and enter the activation code and your account details. The program checks the hardware, selects the appropriate path and sets up the model locally.

Yes. Your model keeps working without us; there is no technical dependency. As your body of knowledge grows, commission a further training and use our program to load the result onto your hardware again. With every model we also hand over the LoRA adapter, the file that carries your specialist knowledge. It is the basis for a later update.

Prepared, but not yet orderable. The route is settled: if your adapter fits the newer base model, it is carried over and trained further on your new material. If it does not fit, your existing model serves as the source the new one learns from. Either way a fraction of the original volume is enough. What is still missing is the orderable process and our own measurement of the resulting quality. We will name a date and a price once both are in place.

Deliberately not. Your model learns your company's stable knowledge but does not assert day-to-day figures such as prices, discounts, stock levels or delivery times, because these change constantly. For such questions it answers helpfully and points to the reliable source, for example your current price list or your sales team. This keeps its answers correct even as figures change.

German and English by default. Every model is trained bilingually so that it answers requests in both languages in your company's style. The terminology and tone come from your documents.