Aleph Alpha Releases Kolibri, Open-Weight AI Model Trained in Germany and Finland
Aleph Alpha has released Kolibri, a 78-billion-parameter German-English mixture-of-experts model trained entirely in Germany and Finland, with weights available under the Apache 2.0 license.
Aleph Alpha has released Kolibri, an open-weight German-English language model that the Heidelberg-based company presents as evidence that advanced artificial intelligence can be developed within Europe rather than imported from American or Chinese providers. The model is available for free download under the Apache 2.0 license, allowing researchers, public institutions and companies to inspect, modify and deploy it without licensing restrictions.
Kolibri is a mixture-of-experts model with 78 billion parameters in total, of which roughly three billion are active for each token processed. This architecture keeps inference costs far below what the total parameter count would suggest, because only a small fraction of the network is engaged for any given piece of text. The model handles both German and English, and more than 21 percent of its training data is German — an unusually high share for a model of this size and a deliberate choice for a company whose home market is German-speaking.
The training run took place on 768 Nvidia B200 GPUs located in Germany and Finland. Aleph Alpha has emphasized the geographic footprint of the compute as a central part of the release, arguing that European institutions need AI systems whose development, data handling and legal framework remain within European jurisdiction. For public administration, healthcare, finance and defense applications, the question of where a model was trained and under which rules can matter as much as its benchmark scores.
The open-weight approach sets Kolibri apart from most frontier models, which are typically accessible only through paid APIs or restricted licenses. By publishing the weights under Apache 2.0, Aleph Alpha gives European universities, startups and government agencies the ability to run the model on their own infrastructure, fine-tune it for local languages and sector-specific tasks, and audit its behavior. That is a meaningful difference for organizations that cannot send sensitive data to external servers.
Aleph Alpha has positioned itself as a sovereign alternative to the dominant AI ecosystems in the United States and China. The company has long argued that Europe's dependence on foreign model providers creates strategic, economic and regulatory risks, particularly as AI systems become embedded in critical public services. Kolibri is the clearest product-level expression of that argument so far: a capable model, trained on European soil, released under terms that maximize local control.
The release also reflects a broader shift in the open-weight landscape. While the largest American labs have kept their flagship models closed, a growing number of European and Asian developers have shown that competitive performance can be achieved with openly available weights. For European policymakers debating AI regulation, public procurement and digital sovereignty, Kolibri offers a concrete example of what a domestic AI stack might look like.
Whether Kolibri can match the capabilities of leading closed models remains an open question, and Aleph Alpha has not framed the release as a direct challenge to the largest American systems. Instead, the company is making a case about control: who trains the model, where the data resides, which laws apply and who can inspect the result. For institutions that answer to European regulators and voters, those questions are increasingly difficult to separate from technical performance.
The model is available now, and its German-language strength is likely to make it particularly relevant for public-sector and industrial users in Germany, Austria and Switzerland. Its success will depend on whether European organizations treat sovereignty as a procurement criterion — and whether an open-weight model trained on European soil can build enough of a developer community to improve over time.
3
