Kolibri releases English-German model weights under Apache 2.0

The researchers behind Kolibri released the English-German language model's weights under the Apache 2.0 licence, according to their technical report. The model has 78.1B total parameters and activates 3.46B per token, using a mixture-of-experts architecture that selects a subset of its parameters for each token. The team trained Kolibri on 24T tokens, with German accounting for more than 20% of the data mix by bytes. Its intended uses include public administration, industry and aerospace. The authors report a competitive balance of quality and serving cost in English and German, but that comparison covers only the open models evaluated in their report.