Aleph Alpha has released Kolibri, an open-weight German-English language model with 78 billion total parameters and roughly three billion active for each token. The mixture-of-experts design routes work through only part of the network, aiming to reduce inference cost while retaining a larger model’s capacity.
German material makes up 21.3% of the training data, supported by a dedicated data pipeline. The company says Chinese models helped generate synthetic training examples. Kolibri was trained on 768 Nvidia B200 GPUs in Germany and Finland and supports context windows of up to one million tokens, according to its technical report.
The weights are available on Hugging Face under the Apache 2.0 license. Aleph Alpha positions the model for public administration, aviation and industrial deployments that value European hosting, legal alignment and stronger German performance.
Company benchmarks put Kolibri at 71% across German evaluations and claim faster decoding than several comparable mixture-of-experts models. Those comparisons include some older systems and should be independently reproduced across real workloads. Open weights improve inspectability and deployment choice, but they do not by themselves establish data quality, safety, factual accuracy or compliance for a specific application.