THKI

In brief: What is THKI?

THKI is TH Köln’s initiative to provide access to AI models for all members of the university community – including lecturers, research staff, students and administrative staff. It is free of charge, complies with data protection regulations and is based on the state-hosted service KI:connect.nrw as part of the KI:edu.nrw project.

THKI offers to ways of access:

• THKI Chat – the models are available via a secure chat interface (web interface). Here, you can submit queries to the latest AI models, upload files and interact in dialogue with the system.
• THKI API – access beyond the interface: You can generate the required API key directly in THKI Chat. Then you can use this key to access the same models in other environments – such as programming environments, scripts or your own applications. This makes the infrastructure more widely usable.

FACTS

Link: https://ki.th-koeln.de
Login: with campusID
Platform: KI:connect.nrw, models integrated via GWDG and OSKI.nrw

TECHNICAL REQUIREMENTS

– TH Köln network, otherwise VPN-connection
– Laptop, smartphone or tablet
– Web browser, no app required

THKI upgrade

Why different AI models?

THKI deliberately offers not just one model, but a selection. Models differ in terms of speed, depth, capabilities, autonomy and cost – and provide different answers to the same question. So, there isn’t just one ‘best’ model, but rather the one that’s right for each situation. By switching between and comparing models, you develop a key AI skill – the critical evaluation of results (referred to as ‘Discernment’ in the AI Fluency Framework). It is only through comparison that blind spots become apparent. Where answers differ, it is worth taking a second look – for example, at hallucinations or one-sided accounts.

You can read about the decision to implement more models and how you can make targeted use of them in the accompanying article „The Model is the Message“.

THKI Chat

Which models can be chosen?

The portfolio brings together models from three infrastructures. The source determines where processing takes place, the level of control over the model’s operation, and whether any costs are incurred.

OSKI.nrw (via RAMSES) – Sovereign inference on the high-performance computer RAMSES in Cologne (University of Cologne), unlimited use:

• GPT OSS 120B (with reasoning, mathematical and logical tasks)

• Mistral Small 4 119B (open high-performance model)

GWDG – proprietary (OpenAI via Azure) – the most powerful commercial models, with usage limits:

• GPT- 5.5 (for demanding tasks)

• GPT- 5.4 mini (small and fast model)

GWDG – Open Source (via KISSKI) – open models, hosted in Germany, unlimited use:

• Gemma 3 27B (efficient open model)

• GLM- 4.7 (for coding)

• Apertus 70B (fully open model)

• Qwen 3.5 397B (biggest open model)

• Qwen 3 Omni 30B (multimodal model)

The models have been build by different developers:
OpenAI (GPT models), Google (Gemma), Z.ai (GLM), Swiss AI (Apertus), Alibaba Cloud (Qwen) and Mistral AI.

Step by Step

Range of functions

• Text generation and text comprehension
• Switch between multiple models
• Editable prompts and searchable chat history
• File upload (PDFs, images, audio files depending on the model) and image and document processing (Vision/OCR)
• Multilingual capabilities
• Output formatting with Markdown and syntax highlighting
  

Use cases

THKI API

With THKI API, you can use the same models outside the web interface – for example, in a programming environment such as VS Code, in your own scripts or applications. This allows you to integrate AI directly into your own work and research context and make wider use of the infrastructure.

The process in brief: You create your personal API key in the THKI Chat (web interface) and then integrate it in the environment of your choice.

Virtual API keys are managed centrally in KI:connect.nrw, where they grant access to the connected models – whilst control remains with the university. The same limits apply as in the web interface.

Note: API key management is currently available to all university staff on a trial basis. This feature is not currently available to students.

• Create & manage: in the user menu under ‘API Key Management’ („API-Schlüsselverwaltung“)

• Endpoint: https://chat.kiconnect.nrw/api/v1

• Guide: Virtual API keys in KI:connect

Step by Step: Create a virtual API Key

1. Log in to THKI Chat: https://ki.th-koeln.de (log in with your campusID).
2. Open the user menu in the bottom left-hand corner (click on your name).
3. Select ‘API Key Management’ („API-Schlüsselverwaltung“).
4. In the dialogue box, select the appropriate user group if necessary and assign a key name.
5. Click on ‘Create key’ („Schlüssel erstellen“).
6. Copy the generated key and keep it in a safe place.
7. In your application, enter the endpoint https://chat.kiconnect.nrw/api/v1 and the key – the interface is OpenAI-compatible.
8. Please refer to the mode overview for the relevant model’s name (API Model Name).

You can view and delete existing keys at any time in the same view. The same limits apply as in the web interface.

Embeddings for your own applications: The API also provides access to the Qwen3 Embedding 8B and E5 Mistral 7B Instruct embedding models (OSKI.nrw). These convert text into vectors and are suitable for semantic search, classification or your own knowledge base (RAG) – using your own data, without the need to operate a central RAG.

Links & literature

Header image: © Zentrum für Lehrentwicklung (ZLE)