Heidelberg-based AI company Aleph Alpha introduced its new language model, Kolibri, on October 3, 2026. Released under the permissive Apache 2.0 license on Hugging Face, the system makes its model weights publicly accessible for enterprise customization. With this release, the German company continues its strategy of delivering transparent, auditable foundational models for organizations that face regulatory constraints regarding proprietary cloud infrastructure.
From an architectural perspective, Kolibri relies on a Mixture-of-Experts (MoE) design. While the model contains 78 billion parameters in total, it activates only roughly 3.5 billion parameters per token during inference. This sparse activation significantly reduces compute requirements during execution, making self-hosted deployments on corporate hardware far more cost-effective. Kolibri is optimized for both German and English, offering support for context windows up to 1 million tokens.
The model incorporates native reasoning capabilities alongside built-in tool-calling functionality. This design allows technical teams to connect Kolibri directly to internal systems, enabling dynamic database queries and workflow automation without fragile external wrappers. Its large context window also facilitates the holistic analysis of voluminous contracts, technical specifications, and legal archives without requiring extensive pre-segmentation.
Aleph Alpha is primarily positioning Kolibri for sovereign industrial and public administration use cases across the DACH region. Many public agencies and medium-sized enterprises handle sensitive records that cannot legally or policy-wise be routed through external cloud services outside the European Union. Because Kolibri can be run fully on-premises or within a private cloud, confidential data never leaves the organization's verified operational perimeter.
By adopting the Apache 2.0 license, Aleph Alpha grants developers full commercial freedom to adapt, fine-tune, and deploy the model within custom applications without licensing fees. Given the relatively low active parameter count, organizations can run sophisticated agentic and reasoning tasks on manageable hardware budgets. The release represents a concrete step toward sovereign, production-grade enterprise generative AI that meets strict privacy mandates.