Journalism begins where hype ends

,,

The danger of AI is not that it will become conscious and hate us, but that it will become competent and ignore us."

—Eliezer Yudkowsky

Aleph Alpha Launches Kolibri-1, an Open-Weight AI Model for Sovereign European Deployment

German AI company Aleph Alpha has released Kolibri-1, a 78.1 billion-parameter open-weight Mixture-of-Experts model with 3.46 billion active parameters per token, targeting sovereign AI deployment in regulated European sectors.
Aleph Alpha Kolibri-1 logo displayed against the German flag
October 5, 2026 12:10 PM IST | Written by Vaibhav Jha

German AI Company Aleph Alpha released Kolibri-1, an open-weight English–German Mixture of Experts (MoE) decoder-only transformer, designed for sovereign deployment in regulated European environments.

Kolibri-1 was released on Hugging Face under the Apache 2.0 license on October 3 as a MoE transformer with 78.1 billion total parameters of which only 3.46 billion (4.4%) activated for any given token.

Aleph Alpha claims this sparsity, combined with hybrid attention, enables competitive quality at substantially lower inference cost than dense models of similar capability.

Trained on 24 trillion tokens with German accounting for more than 20% of data mix, Aleph Alpha claims Kolibri- 1 excels in benchmarks for agentic, maths, code, industry and general knowledge.


“We built Kolibri for sovereign and specialized deployment, with a particular focus on German and on regulated domains such as public administration, industry, and aerospace.
To support these settings, we control the full model-development process, including data, architecture, training infrastructure, post-training, and evaluation,” read a statement from Aleph Alpha.

The German AI company claimed that Kolibri has been designed as per the regulations of newly introduced EU AI Act and its post-training combines supervised fine-tuning and reinforcement learning on reasoning, coding, tool-use, and retrieval tasks, and exposes four discrete reasoning-effort settings so users can trade accuracy against latency.

“Our data pipeline filters against illegal, harmful, and pirated content and redacts personal data from data sources. In addition, we align the answers of the model with the values of the democratic consensus in Europe and train the model to abstain when the provided context does not support an answer,” read an excerpt from the technical paper of Aleph Alpha.

The model is fully available under Apache 2.0 on Hugging Face, allowing users to run it locally on their own hardware, with a detailed tech report covering its training on 24T tokens and hybrid sliding-window attention architecture.

Also Read: ALL IN 2026: Canada and Germany Invest Up to CAD $300 Million in LawZero Amid AI Safety Tensions

Author

  • Vaibhav Jha, editor and co-founder at AI FrontPage

    Vaibhav Jha is an Editor and Co-founder of AI FrontPage. In his decade long career in journalism, Vaibhav has reported for publications including The Indian Express, Hindustan Times, and The New York Times, covering the intersection of technology, policy, and society. Outside work, he’s usually trying to persuade people to watch Anurag Kashyap films.

    LinkedIn

Related Posts

LATEST NEWS