Large language models are AI systems trained on vast datasets to provide human-like responses to text inputs. They are used in areas such as text generation, translation and data analysis.

With Chat AI, you can use various LLMs to, for example, generate text whilst interacting with the system. In addition to open-source models such as LLaMa or Mistral, commercial models such as ChatGPT from OpenAI are also available there.

A key difference between open-source and commercial models lies in their availability and usage. Open-source models are freely accessible and can be customised, but ‘open’ does not mean that everything is transparent when it comes to large language models. Whilst the source code is available for inspection, the training data – which is often crucial – usually remains inaccessible. Rather than complete openness, such models therefore often merely offer the option of free use. Commercial tools, on the other hand, usually offer better performance and additional features, but come at a cost.

The GWDG provides a number of Open Weight models on its own hardware. These models do not pass on any personal data from users to third parties. Inputs (prompts) are stored only for the duration of the session and only locally on the users’ devices.

Commercial models such as OpenAI’s ChatGPT are provided by third parties (in this case, Microsoft). Microsoft reserves the right to store prompts and prompt conversations for up to 30 days. The content of the prompts is therefore transmitted to Microsoft. The requests are sent to Microsoft via GWDG’s servers. Consequently, the request itself does not result in any personal data of the users being passed on.

Please note that the use of commercial OpenAI models is billed according to the consumption of so-called tokens (words or sentence fragments) and is therefore subject to a charge for the UDE. It is therefore worth comparing and using the Open Weight models, as they are free of charge and highly effective.

Tokens refer to the smallest units of an LLM used to process or generate text. Depending on the LLM, a token may comprise individual umlauts or sequences of letters.

Chat AI is part of the Academic Cloud’s range of services. UDE members can log in via Shibboleth using federated authentication.

The following information is partly translated from the GWDG website.

The web interface offers the following functions:

Attach text files (+ button): Add files (.txt) as context for your input.

• Speak (microphone): Speak directly to the selected model. Please note that you must enable microphone access in your browser settings. Following relevant testing, we would like to point out that this function may not work reliably.
• Import/Export: Import or export previous conversations.
• System prompt: Defines the model’s role. See System prompt for more information.
• Generation options: Settings for fine-tuning the model’s responses, temp and top_p. See Generation options for details.
• Share button: Creates a URL with your current settings (excluding conversation content).
• Delete button: Removes all saved data and conversations.
• Memory function: The system’s ability to store information from previous interactions and retrieve it when needed.
• ‘None’: Disables the memory function – each conversation is treated independently.
• ‘Recall’: Adds memory context to the system prompt, allowing the LLM to refer to previous parts of your conversation.
• “Learn”: The system receives automatic memory updates. This feature generates the most natural conversational experience possible.

The interface also provides access to the terms of use, FAQs and language selection (German/English). Users can switch between light and dark modes.

A system prompt is an input that defines the model’s role or behaviour for the current session. It is, so to speak, the task assigned to the language model. By setting a system prompt, you can control how the model responds to your inputs. For example, you can instruct the model to act as a teacher, translator or programmer.

We recommend setting a system prompt before each session to define the desired behaviour of the model.

The system prompt can only be customised for Open Weight models (not for OpenAI models).

Both the DeepSeek models and OpenAI’s o1 and o1-mini possess what are known as reasoning capabilities. This functionality enables the models to tackle more complex tasks that require step-by-step analysis. Examples include multi-step calculations, data analysis or drawing conclusions. Models with this functionality are marked with a special symbol in the model list on the right. Please note that reasoning processes require more processing time, which is why the answer to these tasks is not immediately visible and takes a little longer. Some reasoning models are not capable of streaming content in real time.

Although the available models cannot yet generate images, some of them do offer image analysis capabilities. For example, they can evaluate uploaded images to analyse content such as charts or statistics. Models with image processing capabilities are marked with a camera icon in the model list.

Two key settings influence a model’s responses:

• Temperature (0–2): Controls creativity. Lower values produce predictable responses, whilst higher values encourage creativity.
• Top_p (0–1): Determines how many possible responses are included, or the degree of variability in the cumulative probabilities. Higher values therefore increase the diversity of the generated text. For precise tasks such as programming, a low temperature and Top_p are recommended. Creative tasks benefit from higher values.

These options can only be adjusted for Open Weight models (not for OpenAI models).

GWDG is constantly expanding and updating its range of available models. You can find a current overview of the available models on GWDG’s Chat AI support page. 

GWDG: Overview of available models

The DeepSeek models are AI language models from China that have been developed for general-purpose tasks. They offer good performance and support reasoning capabilities, making them useful for complex tasks. However, it has been observed that these models subject the text they generate to political censorship when queries relate to the Chinese government.

To make the models more efficient, DeepSeek uses a process known as distillation. This involves creating smaller, optimised models from larger ones. These retain much of their reasoning and computational capacity whilst requiring fewer resources. One example is the DeepSeek R1 Distill LLaMa 70B model.

Coding models such as Qwen 2.5 Coder and Codestral 22B are specifically trained to assist with programming tasks. They can generate code, analyse existing programmes, and detect and fix errors. These models are useful for developers.