AG: Generative KI

AI-ToolLab: LLM API Access

Overview

Members of Goethe University can obtain OpenAI-compatible API access to various Large Language Models (LLMs) via the AI-ToolLab.

This service enables experimental access to AI technologies in a protected environment and is provided by studiumdigitale.

Features

  • OpenAI-compatible API:
    Easy integration into existing applications
  • Free use:
    For Goethe University employees (under fair use conditions; restrictions may apply to paid models)
  • Access to 60+ models:
    • various sizes, providers, and hosting locations
    • Language models (LLMs) such as OpenAI GPT-5, Llama 3.3, Mistral or Qwen
    • Code models (Code-LLMs) such as Qwen3 Coder, GLM, CodeLlama
    • Various embedding models for the development of RAG systems

🔑 Requesting API Access

To gain access to the LLM API, you must complete the following steps

  1. Meet the requirements:
    • You must:
    • If you are already using a service of the AI ToolLab, you have already completed this step.
  2. Submit a request:
  3. Account creation:
    • After successful verification, we will create your API account
    • You will receive an email with your individual access data

Notes on the
Training on the Critical Use of AI

  • You can find the training here 📖
  • Log in with:
    • follow the instructions for login to log in with your HRZ account or create a new account.
  • Enroll in the course
  • To complete the course, you must successfully complete all learning units
  • Once you have completed the course, you will receive a badge on Moodle confirming completion
Privacy policy and terms of use

By applying for an API account, you agree to the following measures:

  • Account creation:
    • We will create an account for you in the “LiteLLM” platform hosted by us at GU.
      • For this purpose, we store your HRZ username and email address in the platform and link your account to this data.
      • This is done for the purpose of recognition, any necessary contact, and, if applicable, for billing purposes
  • API usage:
    • When using the API, statistical data is collected about each API call.
      • This includes the exact API endpoint, the time of the request, the input and output token amount, and whether the request was successful or triggered an error.
      • This is done for error analysis and, in the future, possibly for billing purposes.
    • At no time do we store the prompts, or other data that you transmit within the requests.
  • Billing:
    • Currently, the offer is completely free of charge. You don’t have to worry about any costs and can freely test the service in this introductory phase.
    • You start with a €50 budget, and when this is used up, your access to the API will be restricted, so you don’t have to worry about any costs.
    • See “Note: Budget for the use of paid models” for more information

Further Information about the Offer

General Information

Endpoints

We use LiteLLM as an API gateway to provide an OpenAI-compatible interface.
The endpoints are as follows:

Azure-compatible:
https://litellm.s.studiumdigitale.uni-frankfurt.de/

OpenAI-compatible:
https://litellm.s.studiumdigitale.uni-frankfurt.de/v1/

Authentication

For API access, you need an individual API key, which will be provided to you after the request.

You can use this key in combination with the endpoint URL in your API requests.

Available Models

The API currently provides access to 65 different models from three hosting categories.

In principle, the offer includes a mix of commercial and open-source models, optimized in various sizes and for different use cases.

The following models are currently available:

🇪🇺 Azure OpenAI (EU Data Zone)

Hosting: Microsoft Azure in the EU
Costs: Paid (budget-based)

The GU uses the Azure OpenAI Service in the EU Data Zone to provide access to commercial models from OpenAI. These models are generally powerful and offer a wide range of functions.

Models:

gpt-5
gpt-5-chat
gpt-5-mini
gpt-5-nano		GPT-5 model series for demanding or simple, yet very fast execution
gpt-4.1
gpt-4.1-mini
gpt-4.1-nano		GPT-4.1 model series
gpt-4o
gpt-4o-mini		GPT-4o model series
o3
o3-mini		o3 model series
codex-miniCoding model
text-embedding-3-large
text-embedding-3-small		Embedding models
🇩🇪 GWDG/KissKI

Hosting: in Germany at the GWDG via the KissKI project and their service chat-ai.academiccloud.de
Costs: Free under fair use conditions

Through the cooperation with GWDG and KissKI, we can offer a variety of open-source and commercial models. These models are hosted locally in Germany and offer high availability.
We have no influence on the availability and number of models, as these are provided by the GWDG.
If a model is not available, we cannot guarantee that it will be available again.

Overview of the GWDG models: GWDG website

Models:

openai-gpt-oss-120bRecommended: Open source model from OpenAI
mistral-large-3-675b-instruct-2512Recommended: large new Mistral model
gemma-3-27b-it
medgemma-27b-it		Google’s Gemma 3, including a model adapted for medical scenarios
glm-4.7
devstral-2-123b-instruct-2512
qwen2.5-coder-32b-instruct		Coding models
internvl3.5-30b-a3bVision model
llama-3.1-sauerkrautlm-70b-instruct
llama-3.3-70b-instruct
meta-llama-3.1-8b-instruct		Llama models
qwen2.5-vl-72b-instruct
qwen3-235b-a22b
qwen3-30b-a3b-instruct-2507
qwen3-30b-a3b-thinking-2507
qwen3-32b
qwen3-coder-30b-a3b-instruct
qwen3-omni-30b-a3b-instruct
qwen3-vl-30b-a3b-instruct
qwq-32b		Various models from the Qwen2 and Qwen3 series with different focuses (reasoning, thinking, vision, etc.)
teuken-7b-instruct-research
deepseek-r1-distill-llama-70b
apertus-70b-instruct-2509		Other available models

🏛️ Goethe University / studiumdigitale

Hosting: Locally on a server at studiumdigitale at the GU
Costs: Free

Chat models:

llama3.1:8bLatest Llama version
llama3:8bProven all-purpose model
llama2:7b Small text model
mistral:7bCompact language model
codellama:7bCode-specialized
ollama_defaultStandard model (Llama 3.1 8B)

Embedding models:

all-minilm:33m
bge-large:335m
bge-m3:567m
granite-embedding:278m
mxbai-embed-large:335m
nomic-embed-text:v1.5
paraphrase-multilingual:278m
snowflake-arctic-embed2:568m
snowflake-arctic-embed:335m

We update the list depending on availability, demand, and our capabilities.

Instructions: Retrieving current model information

Since the available models change regularly, it is recommended to retrieve the current information directly via the API.

For the following queries, you must have a basic technical knowledge of how to use APIs

/models – Short model overview

Content: Simple list of all available model IDs
Usage: Quick overview of which models are available

Request via CURL:

curl --request GET \
--url http://litellm.s.studiumdigitale.uni-frankfurt.de/models \
--header 'x-litellm-api-key: ihr-api-key'

Example response:

{
  "data": [
    {
      "id": "llama3:8b",
      "object": "model",
      "created": 1677610602,
      "owned_by": "openai"
    },
    {
      "id": "gpt-4o",
      "object": "model",
      "created": 1677610602,
      "owned_by": "openai"
    }
  ]
}
/model/info – Detailed model information

Content: Complete information on all models including hosting and description
Usage: Decision support for model selection

Request via CURL:

curl --request GET \
  --url http://litellm-dev.s.studiumdigitale.uni-frankfurt.de/model/info \
--header 'x-litellm-api-key: ihr-api-key'

Example response (abbreviated):

{
  "data": [
    {
      "model_name": "meta-llama-3.1-8b-rag",
      "model_info": {
        "description": "Meta Llama 3.1 8B RAG",
        "mode": "chat",
        "hosted_by": "KissKI/GWDG in Göttingen, https://kisski.gwdg.de/",
        "input_cost_per_token": 0,
        "output_cost_per_token": 0,
      }
    }
  ]
}

Important information from the API response:

  • model_name: The exact name for API calls
  • description: Short description of the model
  • hosted_by: Hosting provider and location
  • input_cost_per_token/output_cost_per_token: Cost structure (0 = free)
Note: Hosting locations and data security

⚠️ Important note:
Pay special attention to the hosted_by-Feld, der model/info Route as there are three different hosting locations:

  • Azure OpenAI Service in EU Data Zone – Microsoft Azure (EU)
  • KissKI/GWDG in Göttingen – Germany (GWDG)
  • studiumdigitale, Goethe University Frankfurt – Local (Germany)

Depending on the hosting location, different data protection regulations and security guidelines apply. Choose the model that is suitable for your application according to the sensitivity of your data.

Note: Budget for the use of paid models

A starting budget of €50 is currently provided for each API key issued. The allocation of API keys is still in the testing phase. Further rules for budget usage (e.g., top-ups) and an option to view the remaining budget will follow shortly.

The use of GWDG and GU models is free of charge; only the Azure OpenAI models are subject to a fee.

Usage

You can use the API access like a standard OpenAI API access. For this you need your individual API key and the endpoint.

Usage in OpenAI-compatible tools and WebApps

If you are using an OpenAI-compatible tool, and this tool offers the possibility to configure an API key and an OpenAI Proxy/Server/Endpoint, you can use your individual API key and the endpoint. Try both endpoints (with and without /v1/ at the end) to see which one works.

Make sure that you only use the API key with trusted applications, as this grants access to your LLM API account.

!

Usage in different programming languages

You can use the API in any programming language that supports HTTP requests. Here is an example of using the API in Python with the requests library:

Python with requests-Library 
import requests

url = "https://litellm.s.studiumdigitale.uni-frankfurt.de/v1/chat/completions"
headers = {
    "Authorization": "Bearer IHR_INDIVIDUELLER_API_KEY",
    "Content-Type": "application/json"
}
data = {
    "messages": [{
        "role": "user",
        "content": [{"type": "text", "text": "Ihr Prompt hier"}]
    }],
    "model": "llama3.1:8b",  # Beispiel-Modell
    "temperature": 0.7
}
response = requests.post(url, headers=headers, json=data)
print(response.json())
Python with openai-Library 
from openai import OpenAI

client = OpenAI(
    api_key="IHR_INDIVIDUELLER_API_KEY",
    base_url="https://litellm.s.studiumdigitale.uni-frankfurt.de/v1/"
)

chat_completion = client.chat.completions.create(
    messages=[{
        "role": "user",
        "content": [{"type": "text", "text": "Ihr Prompt hier"}]
    }],
    model="llama3.1:8b",  # Beispiel-Modell
    temperature=0.7
)

print(chat_completion.choices[0].message.content)

with other libraries

For examples of using the API in other programming languages or libraries, please visit the LiteLLM documentation.

Important Notes

⚠️ Data protection: It is not permitted to send personal or sensitive data to the LLMs.

📚 Requirement: Completion of the training on the critical use of AI (according to EU AI-VO)

🎯 Fair-Use: Fair-use restrictions apply to GWDG models – avoid stress tests

Support

If you have any problems or questions, write an email to the AI-ToolLabs team.