AI Models
Control which AI models your organisation can use, set the defaults, and add custom models
Choose the AI models DeutschlandGPT may use and set one default per category. Models live under three tabs: Language Models, Image Generation Models, and Embedding Models. Each category always needs exactly one default.
Managing AI models requires the Manage AI model selection permission.
Language models
Set the default
The dropdown at the top sets the organisation-wide default language model. This is the model pre-selected for everyone when they start a new chat.
Users can switch to any other enabled model per conversation from the model dropdown in the chat view. Disabled models never appear there.
Enable and disable models
Every model in the list has a toggle. Disabling one removes it from the model picker for all users immediately.
If a Custom AI Application (Spezialanwendung) is set to use a specific model and that model gets disabled, the application falls back to the organisation's default language model.
Model tags
| Tag | Meaning |
|---|---|
| Recommended | DeutschlandGPT recommends this model for most use cases |
| Europa / Deutschland | Data processed on European or German infrastructure |
| USA | Data processed outside the EU |
| Not GDPR-compliant | This model does not meet GDPR requirements. Enable it only if your use case permits |
| Custom | A model your organisation added (see below) |
Add a custom model
Click Add custom model to connect a model through a custom API endpoint, such as a self-hosted or privately contracted one. Custom models appear in the list with a Custom tag and a ⋯ menu for editing or removal.
Image generation models
These work exactly like language models: set an organisation default and toggle individual models on or off. The default image generation model applies in Custom AI Applications and anywhere image generation runs without an explicit model choice.
Available models include options hosted in Europe (DALL-E 3, Gemini Flash Image, Imagen 4, FLUX.1, Stable Diffusion XL) alongside US-hosted ones. Enable a model tagged Not GDPR-compliant only after checking your organisation's data-processing requirements.
Embedding models
Embedding models power semantic search across document folders and integrations behind the scenes. Only one embedding model is active at a time.
Embedding models are not interchangeable. Embeddings created by one model are incompatible with another, so changing the active embedding model re-indexes every document. In practice, organisations rarely need to change this setting.