Free
€0for one user
- No credit card
- Hosted in Germany
- For testing, without a DPA
Local AI, set up and ready.
With Private Spark, open models such as gpt-oss-120b, Qwen and Mistral run on an NVIDIA DGX Spark in your building. The search index and documents stay on the device, every answer names its source, and you pay a fixed monthly price.
Try the interface for free on the Free plan first. You order Private Spark online.
What is local AI?
With local AI, the language models run on hardware in your own network. Questions, documents and answers do not go to a model provider on the internet. AI services work differently: they send every request to one of the provider's data centres.
For a business, the model is only one part. For a local AI to answer questions about your company, it needs search across your file stores, access rights per person, source references and an interface for the whole team. Private Spark brings all of this together on one device.
The path of a questionswitch to compare
Every question leaves your network, together with passages from your documents.
Everything the AI needs sits with you.
- Open models, locally
gpt-oss-120b, Qwen and Mistral run on the appliance, with no interface to a provider.
- One appliance in your network
The NVIDIA DGX Spark with 128 GB of unified memory needs a power socket and a network cable.
- Search with source references
Every answer names document and page. The search runs across your documents.
- Rights per workspace
Every query checks what the person may see before the model receives a passage.
- Fixed monthly price
Local requests cost nothing extra. Hardware, setup and workshop are included.
- Replacement in 5 working days
If the appliance fails, you receive a replacement device within five working days.
Cloud AI, Private Spark or self-built?
Where your documents are stored, who runs the models and how you are billed.
| Feature | Cloud AI service | Private Spark | Self-built with open models |
|---|---|---|---|
| Documents stay in your building | – | ||
| Models run on your hardware | – | ||
| Delivered set up and ready | not applicable | – | |
| Answers with source references and rights | depends on provider | in-house development | |
| Billing | per user or token | fixed monthly price | hardware and staff |
| Updates as tested packages | by the provider | after your approval | in-house |
| Replacement device if it fails | not applicable | in 5 working days | – |
The columns describe operating models, not individual providers. Private Spark has a management connection to souveraen.ai. As of October 2026.
Run a local LLM yourself?
With some experience you can start open models on your own hardware. The work begins after that: reading in documents from SharePoint, file servers and mailboxes, carrying over rights, backing answers with sources, testing updates and running it all for many users.
Private Spark saves you this self-build. Hardware, platform and models arrive set up and ready, and a four-hour workshop for your team is included. Updates and new models come as tested packages and replace the old version only after your approval.
What a local LLM needs in operationtask by task
- Buy and set up hardwareYour job
- Run and maintain modelsYour job
- Ingest documentsYour job
- Carry over permissionsYour job
- Back answers with sourcesYour job
- Review updatesYour job
- Onboard the teamYour job
- Cover outagesYour job
8 tasks for your team
Local AI with your own data.
The keyword index, vector index and knowledge graph sit on the same device as the models. Every query checks the person's access rights before the model sees a passage. Restricted documents never appear as a citation.
The answer names the document and page from your own records. If you want, you can add larger models from external providers, at their token costs. Your documents stay local even then.
Permissions on every requestPurchasing and HR workspaces
What notice period applies?
- Rahmenvertrag_Lieferant.pdfPurchasing
- Arbeitsvertrag_Muster.docxlocked, not used
The framework agreement can be terminated with three months' notice to the end of a quarter.
Rahmenvertrag_Lieferant.pdf · § 14 · page 6Local does not have to mean offline.
Private Spark has a management connection to souveraen.ai. Models, search index and documents still stay on the appliance. Air-Gap Enterprise works with no outside connection at all. We design it with you based on a quote.
One DGX Spark is sized for a department. When several people work at the same time, they share its capacity. We clarify with you beforehand whether one device is enough or Air-Gap Enterprise fits better.
Data centres in Germany. Open models in hosting, external models without storage. Free, Pro and Max.
Models, search index and documents on an NVIDIA DGX Spark in your network, with a management connection to souveraen.ai.
No connection to the outside at all. We size hardware and models to your workload.
What it costs.
All prices net, excluding VAT. Compare all plans
Pro
€24.90per user per month
- As many users as you need
- With a DPA under Art. 28 GDPR
- Hosted in Germany
Private Spark
€2,790per month, hardware included
- The AI runs on your own premises
- Local models with no token costs
- Can be ordered online
Related topics.
Questions before you start.
Rather talk in person?
We will show you souveraen.ai in 30 minutes, using your own example.
+49 3744 365 2202Which models run locally?
Open models such as Qwen3.6, gpt-oss-120b or Mistral Small 4. Which model answers your questions depends on your tasks and the number of concurrent users, and is set out in the quote.
What hardware does the local AI run on?
On an NVIDIA DGX Spark with 128 GB of unified memory. The device is included in the monthly price and sits in your server room or under a desk.
How much does Private Spark cost?
€2,790 net per month including the NVIDIA DGX Spark, setup and workshop, on a 24-month term.
Does local also mean without internet?
With Private Spark, models, search index and documents sit on the appliance. A connection to souveraen.ai exists for management. Air-Gap Enterprise works with no outside connection at all, based on a quote.
Can we add external models?
Yes, optionally and at the token costs of the provider in question. Your documents stay local even then.
How do updates reach the device?
As tested packages. A new version replaces the old one only after your approval. Nothing is loaded automatically.
What happens if the device fails?
You receive a replacement device within five working days. How your data is backed up is set out in the quote.
Is one DGX Spark enough for our team?
One DGX Spark is sized for a department. If you need more, we design the hardware with you as Air-Gap Enterprise.
Bring AI into your building.
Try the interface for free on the Free plan. You order Private Spark online, on fixed terms.
No credit card needed. Rather call: +49 3744 365 2202