Aleph Alpha releases Kolibri-1, a 78B open-weight German–English MoE with 1M-token context under Apache 2.0

Aleph Alpha published Kolibri-1, a mixture-of-experts model with 78.1B total and 3.46B active parameters per token, 384 experts per layer across 50 layers, trained on 20T tokens of German, English and code with a June 2026 knowledge cutoff. It supports a 1,048,576-token maximum context (262K recommended), tool calling and a reasoning mode, ships as FP8 weights of about 78GB that run on two H100s or a single B200, and the weights are Apache 2.0 licensed. The tiny active footprint makes it cheap to serve, giving European teams a sovereign long-context option without depending on US or Chinese providers.