Kolibri-1

Kolibri-1

Reasons and calls tools in German and English, with a context window of more than a million tokens. An open 78B mixture-of-experts model from Aleph Alpha that activates only 3.46B parameters per token.

Made by Aleph-Alpha on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Kolibri-1 is a reasoning model from Aleph Alpha, built as a mixture of experts and focused on German and English. It has an explicit reasoning mode and supports tool calling. Only 3.46B of its 78B parameters work on each token, which keeps the cost of serving low.

Who it suits: Teams that need an open model for German and English work, such as assistants, document processing and agents, and that can host a large model.

Industry | Who uses it | How they use it | Example
Software | Developers | Build coding and agent features on an open model | An assistant that calls tools in a workflow
Public sector | IT teams | Host a model on their own servers for German-language tasks | An internal question-answering service
Legal and finance | Analysts | Process long documents in German or English | Extracting fields from a long contract
Research | ML researchers | Study a mixture-of-experts model with a long context | A test of retrieval over extra-long inputs

If you only need a small model for a laptop, you can skip it, since it needs about 78 GB of memory.

Where to use it:

• Windows, macOS or Linux with Python, for downloading the files
• Python programs that use the transformers library
• Servers with large accelerators, as the card's hardware list describes

Use case | Who | What they get back
Multi-step reasoning | Developers | Worked answers from the explicit reasoning mode
Retrieval-augmented generation | Search and support teams | Answers grounded in documents you supply
Agentic tool calling | Agent builders | Calls to your own tools from the model
Long documents | Analysts | Processing of contexts up to 1,048,576 tokens

How to set it up:

• Python, to run the Hugging Face download tool
• About 78 GB of memory and disk for the FP8 weights
• Hardware from the card's list, such as 2 A100 80 GB cards, 2 H100 SXM5 cards, or one H200, B200 or B300

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Follow the model card to serve the model on suitable hardware.

pip install -U huggingface_hub
hf download Aleph-Alpha/Kolibri-1

How to use it: You send German or English text to the served model, turn on its reasoning mode when a task needs several steps, and let it call tools when you define them. The card recommends contexts of at most 262,144 tokens for speed and hard tasks.

Example: You give the model a long German report and ask for a summary in English with the key figures listed. With reasoning on, it works through the report and returns the summary.

For newcomers and experts: Start by reading the model card to check your hardware, then try short German and English prompts before long documents. Keep early tests well under the recommended context size.

The model is a 50-layer transformer with 384 experts per layer, one shared and six routed, and its attention mixes sliding-window and grouped-query layers at a 4 to 1 ratio. Most attention layers look at nearby text, which keeps long contexts affordable. The trade is memory, since the whole model must be loaded.

Send German or English text
Router picks 6 of 384 experts
Reasoning mode works the task
Answer or tool call comes back

Good to know:

• License: Apache 2.0
• Cost to run: none named, but the card lists hardware such as 2 H100 cards
• Data it reaches: downloads files from Hugging Face, then runs where you host it
• Not for: small machines, since the model needs about 78 GB for its weights
• Not for: languages other than German and English, which are the ones the card lists

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 78.1B
• Weights: about 78.1 GB in F8_E4M3; about 39.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 13,319
• Likes: 1,011
• License: apache-2.0
• Languages: de, en
• Based on: Aleph-Alpha/Kolibri-1-BF16
• Quantized versions: 28
• Fine-tunes: 8
• Tags: transformers, safetensors, kolibri1, text-generation, reasoning, conversational, fp8

Links

• Model: https://huggingface.co/Aleph-Alpha/Kolibri-1
• Files: https://huggingface.co/Aleph-Alpha/Kolibri-1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Aleph-Alpha/Kolibri-1

Then follow the usage in its model card. Nothing in this kit runs by itself.