Aleph Alpha publishes Kolibri-1 weights under Apache 2.0

Aleph Alpha has published weights for Kolibri-1, a German- and English-focused mixture-of-experts language model, on Hugging Face under an Apache 2.0 license, according to the model’s card, which lists an Oct. 3 release.

The model: Kolibri-1 has about 78 billion total parameters and activates about 3.46 billion per token, the card says. It supports an explicit reasoning mode and tool calling. Context length is listed at 1,048,576 tokens, with a recommendation of at most 262,144 for serving efficiency and complex tasks. Knowledge cutoffs on the card are June 18, 2026, for both English and German. The LICENSE file in the repository is the Apache License, Version 2.0.

Hardware and use: The card puts the FP8 weight footprint at about 78GB and lists minimum hardware such as two A100 80GB GPUs or one H200. It says the model is meant for human-reviewed assistants and agent workflows, not unsupervised systems that act without review. Aleph Alpha is a German AI company; the card names Aleph Alpha GmbH as provider and Aleph Alpha Research GmbH as developer. The card also says Aleph Alpha has signed the EU’s code of practice for general-purpose AI models.

— Rowan Keel