Aleph Alpha released Kolibri, an English-German open-weight large language model (LLM), on October 3, 2026. Licensed under Apache 2.0, the model is built as a Mixture of Experts (MoE) architecture with 78.1 billion total parameters, of which approximately 3.5 billion are active per token.
Model ReleasesAleph AlphaKolibri
Aleph Alpha Releases Kolibri, a 78B Parameter Open-Weight MoE Model for German and English
The model is designed with a focus on "sovereignty," which the company defines as being developed by European teams on infrastructure in Germany and Finland, adhering to European and German laws to ensure data privacy and control. Kolibri supports context lengths of up to 1,048,576 tokens and utilizes a sliding-window attention mechanism across 40 of its 50 layers to maintain efficiency during long-context processing.
To improve linguistic precision for German, Kolibri employs a specialized tokenizer using the UniBPE algorithm. This approach allows the model to process German text using 11.2% to 15% fewer tokens than the GPT-5 tokenizer. Additionally, the model is trained to perform reasoning tasks directly in German, achieving high scores on mathematical benchmarks such as AIME 2025 when prompted in German.
To mitigate hallucinations, Kolibri incorporates the Merlin-Arthur protocol, a method designed to train the model to recognize when information is missing from a provided context. In evaluations on the AA-Omniscience test, the model correctly identified when it did not know an answer approximately 44% of the time.
Deploying Kolibri requires significant hardware resources, including at least 78 GB of GPU memory, such as a single NVIDIA H200 or two NVIDIA A100/H100 80GB GPUs. The model weights are available for download on Hugging Face.
Sources
- Show HN: Germany's new sovereign AI model Kolibri （Hacker News Frontpage, 2026-10-03）
- Aleph Alpha