Aleph Alpha releases Kolibri 1 weights for German and English AI
The German company's new mixture-of-experts model has downloadable weights and targets self-hosted work in regulated sectors.
The release
Aleph Alpha announced Kolibri 1 on October 3 and published the model weights on Hugging Face. The repository’s release commit is timestamped 06:13 UTC that day. According to the company, the German-English mixture-of-experts model has 78.1 billion total parameters but uses about 3.46 billion for each token. It is aimed at organizations that want to run AI on infrastructure they control, including public administration and industrial customers.
What the open release includes
The weights and configuration files carry an Apache 2.0 license. Aleph Alpha’s model card explicitly says that license does not extend to other artifacts, including its code and training methods. Running Kolibri also requires the company’s inference package. The model offers selectable reasoning effort and tool calling. Aleph Alpha says it supports up to one million tokens of context, but the card says the native trained length is 262,144 tokens and recommends that shorter range for complex or latency-sensitive tasks.
What remains to test
Aleph Alpha publishes extensive benchmark comparisons and claims gains over its earlier Kolibri Origin model. Those results come from the developer’s own evaluation setup; they are not independent proof of performance in a particular deployment. Buyers can inspect the released weights and technical report, then measure answer quality, memory requirements and serving cost against their own German and English workloads.