Local AI Models: Why 94% of Requests Never Leave the EU
Local AI: Privacy by Design
Every request to ChatGPT, Copilot, or Gemini leaves the EU. For personal data, trade secrets, and regulated industries, that’s a problem.
IIO’s Approach: AI on Its Own Hardware
Intelego runs its own AI Hub on Hetzner servers in Germany:
| Metric | Value |
|---|---|
| Language models available | 11 |
| Requests on own infrastructure | 94% |
| Cost per token | €0 |
| Data residency | Germany / EU |
Which Models Run Locally?
- Qwen 2.5 7B — general purpose, code, documents
- Llama 3.1 8B — analysis, summarisation
- Nomic Embed — semantic search, RAG
- Phi-3 Mini — fast classification
When Is the Cloud Used?
Only for complex tasks (strategic decisions, legal review) and exclusively after Contextual Data Trust Level B anonymisation.
AI cost analysis for your organisation →
Why Local AI Models Are the Safer Choice
If a language model runs on a US server, US law applies — including the CLOUD Act. That means US authorities could theoretically request access to your company data, even if the data originated in Europe.
Local AI models eliminate this risk entirely.
Which Models Are Suitable?
Recommended for SMEs:
| Model | Size | Strength | Hardware |
|---|---|---|---|
| Qwen 2.5 7B | 4.7 GB | General purpose, German | 8 GB VRAM |
| Llama 3.1 8B | 4.9 GB | Reasoning, code | 8 GB VRAM |
| Mistral 7B | 4.1 GB | Fast, efficient | 6 GB VRAM |
| nomic-embed-text | 0.3 GB | Embeddings/RAG | CPU is enough |
For more demanding tasks (contract analysis, complex reasoning), 70B models on dedicated GPU servers are the better choice.
IIO AI Hub: Local Models as a Managed Service
IIO operates the AI Hub as multi-tenant GPU infrastructure on Hetzner servers in Germany and Finland. No data leaves the EU.
Pricing model: a fixed monthly amount per GPU cell, no token-based surprise costs.
Get to know the IIO AI Hub → IIO offerings →
Next step: Configure your package or talk to us.