Aleph Alpha Debuts Kolibri Open-Weight AI Model

Aleph Alpha has officially released Kolibri, a new open-weight German-English language model designed to balance high performance with low operational costs. Built on a mixture-of-experts architecture, the model features 78 billion total parameters, with approximately 3 billion active per token. The weights have been made publicly available on Hugging Face under an Apache 2.0 license, targeting sectors like public administration, aviation, and heavy industry.

To address European regulatory standards, the developers built Kolibri with the EU AI Act in mind. The training process utilized 768 B200 GPUs located in Germany and Finland. To ensure strong bilingual capabilities, German data accounts for 21.3 percent of the training dataset, supported by a custom-built German data pipeline. Additionally, the team leveraged Chinese models to generate synthetic training data for the project.

According to the technical report, Kolibri achieves a 71 percent score on German benchmarks. The model supports an expansive context window of up to one million tokens. Aleph Alpha claims that Kolibri achieves an optimal balance of quality and operating cost in both languages, outperforming comparable architectures. Specifically, it decodes faster than rival models such as GPT OSS A5B, Qwen 3.6 A3B, and Gemma 4 A4B.

For AI practitioners and enterprise developers, Kolibri offers a highly compliant, cost-effective alternative for deploying large language models within Europe. The combination of a massive context window and fast decoding speeds makes it highly suitable for document-heavy workflows in regulated industries. By utilizing an open-weight model trained locally on European infrastructure, organizations can maintain strict data sovereignty while avoiding the high latency and licensing hurdles of proprietary APIs.