Model selection on your use case
We test several open-source models and sizes on your actual tasks. What counts is output quality, response time, and whether the licence permits commercial use.
On-premise LLMsOn-premise LLMs
On-premise LLMs
on infrastructure you control.
We set up open-source language models so that your data stays inside your organisation or the EU. With permissions, audit logs, and controlled updates.
Book a free AI opportunity review
45 minutes, no strings attached. You leave with a written shortlist.
What it is about
On-premise LLMs
Some data should not reach an external model provider, even under zero data retention. Client files, patient records, engineering documents, or grid data are typical examples. For these cases there are capable open-source models, for instance from the Llama, Mistral, or Qwen families, that you can run yourself.
We take care of the software side: selecting and testing the right model, setting up an inference server such as vLLM, connecting it to your applications, and handling access control, logging, and monitoring. It runs on your hardware or with an EU cloud provider of your choice.
We do not supply hardware and we do not operate data centres. We do help you size the setup, so you neither under-buy nor pay for capacity you will never use.
What we build
The building blocks
Sizing and hosting choice
From model size, user numbers, and context length we work out the GPU capacity required. You decide whether to buy your own servers or rent from an EU provider.
Inference server and API
We deploy an inference server such as vLLM in containers and expose a standard interface. Your applications call the model like any API service, just inside your own network.
Access control and audit logs
Sign-in through your existing user directory, permissions per application or department, and logs kept according to your retention rules. You can always see who used the model and for what.
Monitoring and updates
We monitor load, response times, and errors. New model versions are tested against your evaluation set first and then rolled out on a schedule.
Approach
From the first conversation to operations
- 01
Free AI opportunity review
In 45 minutes we work out whether your case needs a self-hosted model or whether an API model with zero data retention will do. Both have trade-offs.
- 02
Pilot with evaluation
We first run the model in a rented EU environment and measure it on your cases. Only once quality is proven does buying your own hardware make sense.
- 03
Fixed-price setup
Installation, integration, permissions, and monitoring are delivered as a package with a fixed price and acceptance criteria. The configuration is documented and yours.
- 04
Operations and model maintenance
We can take on updates, monitoring, and re-evaluation. Alternatively, we train your IT team to run it themselves.
By industry
Where this applies in practice
The industry overview lists concrete use cases. Detailed pages are available in German.
- Machinery and plant engineeringUse cases: Service and after-sales support
- Tax advisory and law firmsUse cases: Pre-sorting receipts and documents · Supporting client communication · Internal knowledge search
- Medical practices and outpatient centresUse cases: Inbox triage for prescriptions, appointments and requests · Documentation drafts for physician approval
- Industry and manufacturingUse cases: Knowledge assistant for work instructions
- Utilities and municipal energy providersUse cases: Customer service assistant for instalments, meter readings and moves · Internal knowledge assistant for regulations
- Insurance brokersUse cases: Drafting advice documentation under IDD
- Construction and architectureUse cases: Standards and project knowledge assistant
- Agriculture and foodUse cases: Checking labels and allergen declarations against recipes · Preparing IFS and QS audits
Related services
What often comes with it
- Adapting AI modelsWe adapt language models to your tasks, through fine-tuning, small dedicated models, or prompting and structured outputs. Measurement decides which route fits.
- RAG knowledge assistantYour staff ask questions in plain language and get answers drawn from your own documents. Every answer points to the passage it came from.
- AI developmentWe build AI solutions that plug into the systems you already run and deliver measurable results. One team from the first question to day-to-day operation.
FAQ
Questions about On-premise LLMs
How much does it cost to run an LLM on-premise?
There are three cost blocks: hardware or rented GPU capacity, setting up the software, and ongoing operations. You buy hardware directly from a reseller or hosting provider. For setup and operations we quote a fixed price per package after the pilot.
What hardware does a local language model need?
It depends on model size, quantisation, context length, and the number of concurrent users. Smaller models often run on a single data centre GPU, larger ones need several. We measure the requirement during the pilot, before you buy anything.
Are open-source LLMs as good as ChatGPT?
For many business tasks, such as summarising, extracting data, or answering questions about your own documents, current open-source models are good enough. For very complex reasoning, the largest commercial models often still lead. We compare both on your cases.
Can we use open-source models commercially?
That is governed by each model licence, and the terms vary. We check the licence during model selection and document it. A final legal assessment should come from your legal team or counsel.
On-premise or EU cloud: which is better?
Your own hardware pays off under sustained high load, or when data must not leave the building at all. An EU cloud is more flexible, available sooner, and almost always the better place for a pilot. Many companies start in the cloud and decide afterwards.
Who handles model updates?
New model versions appear often, but not every one is better for your case. We test updates against your evaluation set and only roll them out when the results hold up. This can be part of an operations contract or handled by your IT team.
On-premise LLMs: let's find out in 45 minutes whether it pays off for you.
Book a free AI opportunity review
45 minutes, no strings attached. You leave with a written shortlist.