Kolibri-1

Kolibri-1 is a language model that reads and writes German and English, follows instructions and calls software tools. Aleph Alpha released it on 3 October 2026 with downloadable weights, allowing organisations to run it on their own equipment. The launch announcement describes a model trained from the start for these two languages.

A small part of the model works on each token

Kolibri-1 contains 78.1 billion parameters, the numerical values learned during training. About 3.46 billion are active for each token, or piece of text. This mixture-of-experts design selects part of the model for each step while retaining the full set of weights in memory.

Users can choose four reasoning settings: none, low, medium and high. Higher settings allow more work before the answer. The model card documents structured tool calls and gives examples for running the model.

The recommended context is 262,144 tokens

Kolibri-1 was trained with sequences up to 262,144 tokens. Its configuration extends the context window to 1,048,576 tokens, but Aleph Alpha recommends staying within the trained length for complex tasks and efficient serving. A context window holds the prompt, supplied material and generated text.

The weights use the Apache 2.0 licence. Running them still needs suitable hardware: the model card lists substantial GPU memory requirements, including two 80 GB A100 or H100 cards for the FP8 version. Organisations evaluating it should test German and English tasks separately and include their own documents and tools.