The Model is the Message
THKI Chat doesn’t offer you just one model, but a whole range to choose from. This isn’t a coincidence – it’s a feature. We’ll show you why it’s worth looking beyond the default model – and how you can use this to achieve better results.
Why several models?
In 1964, Marshall McLuhan coined the phrase ‘The medium is the message’: it is not just the content that matters, but the medium through which it is conveyed. Researchers Benjamin Bratton and Blaise Agüera y Arcas have applied this idea to AI – „The Model is the Message“ (Noema, 2022). Put simply: a language model is not a neutral channel. Training data, architecture and optimisation vary from model to model – and so do the results. Asking two different models the same question leads to two different answers. The chosen model is therefore part of the message.
Which model is useful depends on the tasks at hand. Models differ, for example, in terms of:
- Speed and depth – a small model responds within seconds; a reasoning model involves more intermediate steps but delivers more reliable results.
- Capabilities – such as programming, processing images and documents, or multilingual writing.
- Origin and sovereignty – commercially and externally hosted, or openly hosted on German or North Rhine-Westphalian infrastructure.
- Costs – commercial models charge per query, whilst open models offer unlimited use.
What you get out of it
Choose the right tool for the task, rather than using the same model for everything.
- Cross-check: Ask the same question to several models. Where the answers differ, it’s worth taking a closer look – this will help you spot hallucinations and one-sided accounts, as well as support you in better recognizing the limitations of each model.
- Informed choice: Does the task really require a commercial model – or would the data be better off on a model hosted under your own control?
- Understanding response behaviour: With a bit of practice, you’ll learn how a model reacts to specific prompts and be able to formulate them more effectively.
Quick start
Which model is best suited for what?
- For everyday use: Mistral Small 4 – the default model. A European all-rounder with reasoning capabilities, image understanding and a large context window; it can be used without restriction and is independently hosted at the University of Cologne. For most tasks, you won’t need to switch models at all.
- For Code: GLM-4.7 – specialised in programming, terminal tasks and tool usage.
- For the most demanding tasks: ChatGPT 5.5 – the most powerful model in the catalogue. As a commercial model, it is charged per query and is therefore subject to usage limits.
The full catalogue is listed below – sorted by source, as these determine where your data is processed.
Where your data is processed – the icons in the model menu
Each model in the drop-down menu is marked with an icon indicating where your queries are processed:
- German flag – the open models. These are hosted on servers in Germany: at the GWDG in Göttingen or via OSKI.nrw on NRW infrastructure in Aachen and Cologne. Your data does not leave Germany.
- EU symbol – ChatGPT 5.5. Processing takes place via the GWDG Azure connection on European servers.
- Globe icon – ChatGPT 5.4-mini. Processing takes place via the GWDG Azure connection on global servers.
Regardless of the model: Personal or sensitive data must not be entered. All queries are routed through the central KI:connect interface and are attributed to TH Köln, not to individual users.
The model catalogue in THKI Chat – by origin
In THKI Chat, we bring together models from three infrastructures. The source is not merely a technical detail: it determines where your data flows, how independently the model is operated, and whether any costs are incurred.
OSKI.nrw (via Inferenz NRW)
OSKI.nrw‘s (Open Source-AI.nrw) state-owned, sovereign AI inference, operated on infrastructure in Aachen and Cologne (including the RAMSES supercomputer at the University of Cologne). Unlimited use; the data remains in North Rhine-Westphalia.
- Mistral Small 4 119B (Mistral AI) – the default model in THKI chat. A high-performance European model with a large context window (around 262k) and strong multilingual all-round capabilities, including image recognition, tool usage and reasoning.
- GPT OSS 120B (OpenAI, open) – OpenAI’s open model, optimised for fast inference: strong all-round and reasoning performance, and good for use with tools; around 131k context. A fast, unlimited alternative to the default Mistral Small 4.
GWDG – proprietary (OpenAI via Azure)
The commercial OpenAI models, provided by GWDG via Microsoft Azure. They are among the most powerful models available, but are hosted externally and are subject to usage limits as they are charged per query. The current limit is displayed directly in the model menu. Once it has been reached, simply switch to an open model – these run indefinitely.
- ChatGPT 5.5 – OpenAI’s high-end model for the most demanding tasks: complex specialised work, coding, tool-intensive agents and workflows where execution quality matters. Context window of around 256k tokens. Processed on EU servers.
- ChatGPT 5.4-mini – The fast, cost-effective option for routine queries; around 400k context. Processed on global servers.
GWDG – Open Source (via KISSKI)
Open models, operated by GWDG via its AI service centre, KISSKI, in Germany. Open weights, unlimited use, no charge per query.
- GLM-4.7 (Z.ai) – Coding-focused: strong in multilingual, agent-based programming, terminal tasks, tool usage and reasoning (around 200k context).
- Apertus 70B (Swiss AI) – A fully open-source model from Switzerland that covers over 1,800 languages and is based exclusively on legally sound, open training data – offering maximum transparency and autonomy (around 65k context).
- Qwen 3.5 397B (Alibaba Cloud) – A MoE model (397 billion parameters, 17 billion of which are active) and one of the most powerful open models in existence: outstanding in reasoning, coding and mathematics, with image understanding and around 262k context.
- Qwen 3 Omni 30B (Alibaba Cloud) – Omni-modal: processes text, images, audio and video, including speech recognition and voice dialogue in many languages. For multimodal and speech-based scenarios (around 65,000 contexts).
Background
The context window indicates how much text a model can process simultaneously– including input, uploaded documents and conversation history. This is measured in tokens, which are the models’ processing units; one token roughly corresponds to a short word or part of a word (you can try this out in OpenAI’s Tokenizer). As a rule of thumb: 100k tokens correspond to around 75,000 words or approximately 100–150 pages of text – slightly less for German texts, as they are broken down into more tokens. The larger the window, the longer the documents and conversations can be.
The following applies to all models: training data and optimisation shape the responses – no model is neutral. Examine the results critically.
Models for Embedding
Not every model is designed for chat. Embedding models do not engage in dialogue – they convert text into numerical vectors that make meaning machine-comparable. This enables semantic search, classification or the creation of a dedicated knowledge database (RAG). They can only be accessed via the THKI API, which is available to staff and lecturers – not via the chat window.
Qwen3 Embedding 8B and E5 Mistral 7B Instruct are available, both via OSKI.nrw on NRW infrastructure (in Aachen and Cologne respectively). The advantage is that you can build your own search and RAG applications on your own data – without having to operate a central RAG and without your documents leaving your own secure environment.
Comparing reasoning – a simple test
The quickest way to understand the difference between a direct answer and internal intermediate steps is to compare them: ask two open models – GPT OSS 120B (with reasoning) and Apertus 70B (which answers directly) – a challenging question from your field of expertise. Pay attention to three things: how long does the model take? How transparent is the solution process? And is the result correct?
Try it out
You can change the model in the top-left corner of the model drop-down menu; clicking “Erfahre mehr” (‘Find out more’) will take you to the full model overview. Try them out, compare them, get a feel for them – because at the end of the day, what matters is: The Model is the Message.
→ Visit THKI Chat
Any questions?
Contact digitalelehre@th-koeln.de!
Links & literature
Header image: © Zentrum für Lehrentwicklung (ZLE)
