Your AI runs on our stack.
In Germany and the EU.
SafeChats and our custom AI solutions run on leading open-weight models in European data centres. No calls to US AI providers, no training on your data, and you can see exactly which model answers you.
Open-weight models
EU data centres
No US hyperscaler
The stack, top to bottom
In operation
05
Your applications
SafeChats, custom agents, integrations
04
Private retrieval
Your documents, indexed separately per customer
03
Open-weight models
Mistral · Qwen · DeepSeek · Kimi · Hermes
02
GPUs
NVIDIA H200 · B300 · GB300
01
Data centres
Germany and the EU
Where your data goes.
And where it never goes.
01
You ask
A question or a document from your team, sent over an encrypted connection.
02
Your context
Private retrieval finds the relevant passages in your own documents, separated per customer.
03
An open model answers
An open-weight model runs on GPUs in an EU data centre and writes the answer.
04
Back to you
The answer comes back with its sources. Nothing is kept for training.
Where your data never goes
These are not settings you have to switch on. They are how the stack is built.
✕
No calls to OpenAI, Google, Anthropic or other US AI APIs
✕
No AWS, Azure or Google Cloud in the chain
✕
No training on your prompts or documents
✕
No processing outside the EU
The best open models.
Running in Europe.
Mistral AI · France
Mistral
Strong in European languages and business writing.
Alibaba · China
Qwen
All-rounder for analysis, coding and multilingual work.
DeepSeek · China
DeepSeek
Step-by-step reasoning for maths, logic and complex questions.
Moonshot AI · China
Kimi
Long documents and multi-step agent tasks.
Nous Research · USA
Hermes
Precise instruction following and tool use for agents.
Line-up as of October 2026. Open models improve fast, so this list changes; we keep it current.
Open weights, explained
Where a model was developed is not where your data goes.
Open-weight means the model files are published. We download them and run them on our own GPUs in EU data centres. The model has no connection back to its developer, wherever that developer is based. What decides the safety of your data is where the model runs, and that is always in Europe.
No lock-in.
No black box.
01
No lock-in
Switch models without switching provider or rebuilding your setup.
02
Inspectable
Open models can be tested and evaluated independently, not just trusted.
03
Tailored to you
Fine-tune a model on your domain language and your documents.
04
Always current
New open models arrive every few months. You get the best one without a migration project.
Enterprise-grade security.
Switched on by default.
Encryption everywhere
Data is encrypted in transit (TLS 1.2+) and at rest (AES-256), including your document index and uploads.
Customer isolation
Every customer environment is separated. Your data is never mixed with anyone else’s.
Single sign-on
Connect your identity provider via OIDC or SAML 2.0 (e.g. Microsoft Entra ID, Okta, Google Workspace) and keep your password and MFA policies.
Audit logging
Admins see who used which assistant and when, for accountability and your internal policies.
Three ways to run it.
Same stack, your choice.
Fastest start
Managed
SafeChats and your agents on our managed stack in EU data centres, fully maintained by us.
Setup: days
Capacity: shared GPUs, data separated per customer
Best for: most SMEs
Reserved for you
Dedicated GPUs
Your own GPU capacity on NVIDIA H200, B300 or GB300, rented for you and used by no one else.
Isolation: physically separate hardware
Capacity: sized to your load
Best for: high volume, sensitive data, own fine-tunes
Your data centre
On-premise
We deploy the containerised stack in your own data centre, offline if needed.
Setup: custom project
Connectivity: can run air-gapped
Best for: critical infrastructure, defence, finance
The questions
we get asked most.
Something else? Ask us directly, we are happy to explain the details.
Why do you use models developed in China?
Where exactly does processing happen?
Are open models as good as ChatGPT?
Do you train on our data?
What is the difference between managed and dedicated GPUs?
See it with
your own documents.
Start a SafeChats trial on the managed stack, or talk to us about dedicated GPUs or an on-premise setup.