Aleph Alpha Kolibri-1 opens 78B German English MoE weights

Kolibri is a statement that Europe does not want to be only a rule maker in AI, it wants credible base models of its own. By opening 78.1 billion parameters of German English MoE weights under Apache 2.0, Aleph Alpha gives European governments and regulated industries a serious local option for long document workflows, public administration and industrial knowledge systems. The technical story is not just size, it is design choices that reflect those use cases: sparse activation for serving cost, heavy German pretraining and a context window that can swallow entire statute books.

In the race to AGI, Kolibri matters less as a frontier capability leap and more as a sovereignty milestone. It shows that highly capable, open-weight models are no longer an almost exclusively Chinese phenomenon. Combined with the EU AI Act’s transparency rules and the lab’s detailed training content summary, it also sketches a different path to powerful systems, one where openness and documentation are part of the brand. If more European buyers standardize on Kolibri class models instead of US or Chinese APIs, you get a more diversified ecosystem of powerful models, more independent evaluation and a wider base of actors experimenting at the edge of what these systems can do.