Aleph Alpha releases Kolibri, a sparse sovereign open-weight MoE

Kolibri is reported as an Apache 2.0 open-weight model with about 3.46B active and 78.1B total parameters, targeting German-language, reasoning, coding, agentic, and long-context workloads. Its sparse design may support efficient inference, but total-weight storage, quantized VRAM needs, throughput, and benchmark methodology remain unclear; the release has already circulated in recent content.