Aleph Alpha releases sovereign open-weight Kolibri-1

Aleph Alpha's Kolibri-1 is reported as an Apache-2.0 German sovereign model for government and regulated deployments, using a 78.1B-total/3.46B-active MoE architecture with reasoning modes, on-premise support, and a hosted trial. The advertised one-million-token context conflicts with a reported 262K training length, while benchmark evidence and Qwen comparisons remain thin and informal.