Aleph Alpha Releases Open-Weight Kolibri Model for German and English
Germany’s Aleph Alpha has released the Kolibri model with publicly available weights under the Apache 2.0 license. It is a bilingual Mixture-of-Experts model also intended for regulated environments.
German company Aleph Alpha announced the release of Aleph Alpha Kolibri on October 3, 2026, and made its weights available on the Hugging Face platform. It is a bilingual language model for German and English, designed as a Mixture-of-Experts (MoE) model. The company is primarily targeting organizations that want to operate the model under their own data controls, including public institutions and regulated businesses.
Kolibri has a total of 78.1 billion parameters, but activates approximately 3.46 billion parameters when processing a single token. This MoE arrangement distributes computation among specialized parts of the model instead of activating the entire network at every inference step.
Aleph Alpha Kolibri and the Weight License
The model weights and configuration files are published under the Apache 2.0 license. This allows them to be used and operated independently in accordance with the license terms. However, the model card also explicitly states that the license does not cover other artifacts, code, architecture, or training methods.
Kolibri is therefore more accurately described as an open-weight model, rather than a fully open-source project in the broader sense. Publicly available weights give organizations the option to operate the model on their own infrastructure, but do not disclose the entire technology and training stack.
Aleph Alpha uses the term “sovereign” for Kolibri. This is a company-defined designation based on control over the model’s development, training, and deployment, not an independently certified status. The company says it trained the model in Germany and Finland.
Long Context, Tool Calling, and Reasoning
The model supports tool calling, meaning the invocation of external tools as part of task processing, as well as controlled reasoning. The manufacturer declares a maximum context length of 1,048,576 tokens. For demanding tasks, however, it recommends working with no more than 262,144 context tokens.
Long context may be relevant for tasks involving extensive documents or larger collections of textual source material. The stated capacity alone, however, does not indicate the practical demands on memory, inference speed, or response quality at individual input lengths.
Even with an MoE architecture, Kolibri remains a large model. The availability of its weights does not mean that self-hosting requires little hardware. Aleph Alpha declares hardware requirements, but real-world performance across various optimized implementations and with long context will need to be verified through independent measurements.
What Has Not Yet Been Confirmed
The company has published benchmark comparisons and claims about cost efficiency, but these are materials from Aleph Alpha itself. Independent performance reproductions, security tests, and broader experience from production deployments are not yet available.
For Aleph Alpha Kolibri, it will therefore be especially important to monitor independent benchmarks and measurements of inference costs, the availability of optimized implementations, and actual hardware requirements when working with long context. The first confirmed deployments in public administration and regulated sectors will also be relevant.
Sources
- Aleph Alpha – Kolibri Has Landed: A Sovereign Open-Weight Model – Confirms the release date, public availability of the weights, the Apache 2.0 license, parameters, context, and the company’s claims about training and sovereignty.
- Hugging Face – Aleph-Alpha/Kolibri-1 – Confirms the publicly available model card, technical parameters, release date, limitations, and the precise scope of the license covering the weights and configuration.
- Aleph Alpha – Kolibri – Confirms the model’s public availability, stated hardware requirements, tool-calling support, and the recommended context of 262,000 tokens.
Verified and updated: 10/03/2026 15:26
