Aleph Alpha has released Kolibri-1, an open-weight Mixture-of-Experts language model designed for bilingual English and German tasks. The model features 78.1 billion total parameters but activates only 3.46 billion per token, allowing it to run on a single B200, B300, or H200 GPU.
- Kolibri supports up to 1,048,576 tokens of context using a hybrid attention architecture with sliding-window and full-attention layers.
- It was trained on 24 trillion tokens, including over 2 trillion curated German tokens, followed by supervised fine-tuning and reinforcement learning.
- The model achieves top scores among its competitors in English (75.5) and German (70.8) benchmarks, leading GPQA Diamond with 84.3.
- Weights are available under the Apache 2.0 license on Hugging Face, optimized for sovereign deployment in regulated sectors like public administration.
The release targets compliance with EU regulations such as the AI Act and GDPR, offering a deployable solution for industries requiring data sovereignty and high-performance reasoning.