Aleph Alpha Releases Open-Weight Kolibri with 1M Context Window

Aleph Alpha has released Kolibri, an open-weight Transformer model designed for sovereign, mission-critical applications in government and regulated industries. The English-German Mixture-of-Experts architecture features 78.1 billion total parameters with 3.46 billion active per token, supporting context windows up to one million tokens. Weights are available on Hugging Face under the Apache 2.0 license, enabling on-premises deployment. Trained on 768 B200 GPUs using nearly 24 trillion tokens, Kolibri utilizes 384 experts with six active per token and a hybrid attention mechanism (full attention in 10 layers, sliding window in 40). In vendor-run benchmarks, Kolibri achieved an overall score of 75.5 in English and 70.8 in German, with specific scores of 96.9 on AIME 2025 and 85.9 on LiveCodeCodeBench v6. The model emphasizes grounding, avoiding wrong answers on 44% of AA-Omniscience items compared to 14.8% for its predecessor, Kolibri Origin.