Kolibri is an open-weight LLM from Aleph Alpha for German and English

This model offers German and English language users a privacy-first, sovereign AI deployment option at no cost.

Aleph Alpha has released Kolibri, an open-weight large language model (LLM) designed for German and English. This model, with a total of 78 billion parameters, uniquely utilizes a Mixture of Experts (MoE) architecture, activating only about 3.5 billion parameters per token. Kolibri was trained from scratch on infrastructure in Germany and Finland, with a strong emphasis on EU data privacy and regulatory compliance, particularly the EU AI Act. It was released under the Apache 2.0 license on October 3, 2026, with weights available on Hugging Face. Key features include a specialized tokenizer for German compound words, efficient sliding-window attention for long context handling (up to 1 million tokens), native German reasoning capabilities, and training designed to reduce hallucinations by enabling the model to state when it doesn't know an answer. The model's development prioritizes data sovereignty and intellectual property safety, allowing deployment on private servers.

Kolibri's architecture as a Mixture of Experts (MoE) model, activating only a fraction of its parameters per token, significantly reduces computational cost during inference. This approach is particularly beneficial for organizations concerned with efficiency and resource management, while still requiring substantial memory to hold all parameters. The model's "sovereign" design, developed within Germany and Finland with EU regulations in mind, provides a unique selling proposition for businesses and governments prioritizing data control and compliance, making it an attractive alternative to models developed under different regulatory regimes.

The advanced tokenizer and sliding-window attention mechanisms are key innovations that directly address common LLM limitations. By tokenizing German more efficiently and enabling extremely long context windows with reduced computational overhead, Kolibri offers superior performance for German language tasks and complex document analysis. The focus on training the model to admit when it doesn't know an answer, through the Merlin-Arthur protocol, directly combats the hallucination problem, a critical concern for reliable AI applications, especially in sensitive domains.