The Data Center provides an AI service for JGU. This includes various large language models. The ZDV controls and hosts the service itself, so the data (except for the web search) remains within JGU.
Log in
You will need a JGU account to log in. The AI chat is initially available to employees. The offer for students is currently being prepared.
Log in at https://ki-chat.uni-mainz.de with the login details of your JGU account.
Available AI Models
Model updates each semester
In future, it is planned to update the models once per semester. Where possible, this will take place during the lecture-free period and will be announced in advance.
- The model updates will not result in any fundamental changes for users. The models will become more powerful, more reliable, and of higher quality.
- In addition, the AI chat user interface (OpenWebUI) can also be updated and enhanced with new features during breaks between semesters.
- Predefined prompts that have already been created can normally continue to be used.
Current, discontinued models
| AI Model | Description | Origin/Level of Knowlege | Usage |
| GPT OSS 120B | Highly efficient model, requires less computing power, comparable to o3-mini/o4-mini, optional reasoning* for improved logic capabilities | OpenAI, Release 08/2025, Level of Knowledge 06/2024 |
General chat application, create and summarize texts |
| Qwen3 235B (VL/Thinking) |
Duo of two models with top performance, especially in STEM* and coding applications (with reasoning) or text work (without reasoning) | Alibaba Cloud, Release 07/2025, Level of Knowledge 2024 |
STEM** (Reasoning on) Agentic Coding (on/off) Text work (off) Detailed knowledge (off) Image Processing (off) |
| Qwen2.5 Coder 32B | Specialized in programming tasks, accurate & fast coding at the level of the best models without reasoning | Alibaba Cloud, Release 08/2025, Level of Knowledge 10/2024 |
only API |
*Refers to the ability of the model to create an internal monologue to give more precise or logical answers.
**Science, Technology, Engineering and Mathematics
New models (subject to change)
| AI Model | Description | Origin/Level of Knowlege | Usage |
| MiniMax M3 |
Multimodal Mixture-of-Experts model (428 billion parameters, with approx. 23 billion active) featuring dynamic reasoning* (on/adaptive/off). Official documentation: Quantization used: |
MiniMax, Release 06/2026, Level of knowledge 01/2026 |
General chat application, complex and agent-based tasks, coding, creating and summarizing text |
| Qwen3.6 35B | Official documentation and manufacturer Benchmarks: https://huggingface.co/Qwen/Qwen3.6-35B-A3B Quantization used: https://huggingface.co/Qwen/Qwen3.6-35B-A3B-FP8 |
Alibaba Cloud, Release 04/2026, Level of knowledge not published |
General chat application, coding, creating and summarizing text |
*Refers to the ability of the model to create an internal monologue to give more precise or logical answers.
Image models
| Model | Description | Usage |
| Z-Image-Turbo |
Optimized version of the Z-Image image creation model Official documentation: |
Creating images |
| Qwen-Image-Edit-2511 Lightning Variante | Optimized version of Qwen-Image-Edit-2511Official documentation: https://huggingface.co/lightx2v/Qwen-Image-Edit-2511-Lightning |
Editing Images |
System Components
- Frontend (OpenWebUI): User interface for interaction.
- Load balancer (LiteLLM): Central interface to AI server(s).
- AI server (vllm): Executes the AI models on the GPUs.
Overview of Functions and Applications
News and Plans
All Posts and News combined: https://www.en-zdv.uni-mainz.de/tag/ai/
We are currently working on the following projects:
- Enabling additional features for the AI chat user interface (OpenWebUI)
- Tool-calling integrations (e.g., search in arXiv)
Overview of Frequently Asked Questions (FAQ)
- How long can I save Chats?
- How can I share and export Chats?
- How can I upload Files and what do I need to consider?
- How does voice-based interaction (voice input and audio processing) work?
How do I optomize my Results?
For optimal results, please formulate your query in English. English is clearly dominant in the training data for all AI models.
Data Protection
All components of the system are hosted and operated by the university's Data Center , which means that all data remains exclusively on Data Center servers.
The only exception is the web search function. When web search is activated, the user's prompt is enriched with additional information. This could be the current date, for example. If necessary, the prompt is sent in summary form to our search provider Tavily (https://www.tavily.com/) in order to answer the search query. The prompt is then discarded in accordance with the contract/settings.
As a rule, no files or images are sent along with the prompt. In some cases, however, the model that generates the search query may use relevant information from the context to enrich the query.
Support
Members of the University of Mainz with application questions please contact the Digital Teaching Competence Team.
Email: digitale-lehre@uni-mainz.de
Members of the University of Mainz with technical questions please contact the unix group of ZDV.
Email: unix@uni-mainz.de