Kolibri: A Sovereign Open-Weight Model

You can now download and deploy a high-performance AI model that offers explicit data sovereignty and operates efficiently on-premise, bypassing third-party inference services.

Aleph Alpha has released Kolibri, an open-weight, Mixture-of-Experts Transformer model. Kolibri has 78 billion total parameters, with 3 billion active, and supports context lengths of up to 1 million tokens. The model is available for download on Hugging Face under the Apache 2.0 license. Kolibri was developed using a refined model training pipeline, allowing for rapid iteration. Its predecessor, Kolibri Origin, had 30 billion total parameters and a 65,000 token context window. Kolibri is specialized for mission-critical work in regulated sectors such as public administration, industrials, and aerospace, with an emphasis on German language understanding, reasoning, and agentic capabilities. Aleph Alpha highlights Kolibri's performance against models with significantly more active parameters and its efficiency for on-premise deployment, ensuring data sovereignty for its customers.

Kolibri's release signifies a growing trend towards "sovereign AI" solutions, designed to meet the stringent data privacy and regulatory demands of European enterprises and governments. By offering full supply chain transparency, freedom of deployment, and IP safety, Aleph Alpha addresses a key concern for organizations operating under regulations like the EU AI Act and GDPR. The model's efficient architecture, utilizing 3 billion active parameters out of 78 billion total, is positioned to optimize the balance between capability and deployment cost, a critical factor for mission-critical applications where ROI must be measurable.

The emphasis on a native German and English bilingual design, rather than translation, aims to provide more authentic performance for German-speaking markets. This approach, coupled with rigorous internal evaluation suites mirroring specific industry needs, suggests a market shift towards highly specialized, yet versatile, AI models. Customers can expect future developments to focus on further refining contextualized performance and explainability, particularly for use cases requiring high degrees of trust and compliance.