German company Aleph Alpha has released Kolibri, an open-weights AI model designed for government and regulated industries. The model features 78.1 billion parameters and utilizes a Mixture-of-Experts architecture. It supports a context window of up to 1,048,576 tokens and is available on Hugging Face under the Apache 2.0 license.
The model was trained on 768 Nvidia B200 GPUs using a total of 23.64 trillion tokens across three stages. The training data consisted of 62% English, 21.3% German, and 14% code. Aleph Alpha developed a custom tokenizer to handle the morphology of both German and English, including complex compound words.
Kolibri allows organizations to run the model on their own infrastructure to maintain data sovereignty. It complies with the EU AI Act, the GDPR, and the Code of Practice for general-purpose AI. The model requires at least one Nvidia H200, B200, or B300 GPU, or two A100 80GB/H100 SXM5 units for operation.
The developer reports that Kolibri achieved 96.7% on the AIME 2026 benchmark and 84.8% on GPQA Diamond. Aleph Alpha claims the model performs comparably to competitors with up to four times more active parameters in tasks involving mathematics, programming, and long-context processing.
