Aleph Alpha opens Kolibri’s weights for self-hosted German and English AI

Its Apache 2.0 licence permits use and modification under its terms. The public repository contains model files, rather than a waiting-list announcement. [1][4]

Kolibri separates two easily confused costs. It holds 78.1 billion parameters but uses about 3.46 billion for each token: roughly 4.4%. That reduces the portion involved in a calculation; it does not remove the rest from the memory budget. The active fraction is not an equivalent reduction in electricity, latency or operating cost. [1][2]

Aleph Alpha lists approximately 78 GB for the eight-bit weights and 156 GB for the higher-precision version. Operation needs additional capacity. Its longest trained context is 262,144 tokens; the advertised million-token reach extends beyond that length, and the developer recommends the shorter limit for complex tasks. [2][3]

For a team handling German or English documents, the useful next step is a bounded comparison on its own work: answer quality, unsupported claims, time to completion and hardware cost. Downloadable weights expand deployment choice. They do not establish that self-hosting is the best or cheapest choice.