tl;dr: Aleph Alpha released Kolibri, an Apache 2.0-licensed 78B-parameter (3B active) English-German MoE Transformer with up to 1M token context, targeting sovereign, regulated deployments like German public sector, aerospace, and industrials. The model was trained in Germany/Finland on 768 B200 GPUs over ~24T tokens using a custom bilingual tokenizer, Merlin-Arthur grounding to reduce hallucinations, and curated German data (21% of pretraining). Aleph Alpha claims Kolibri sits on the quality/cost Pareto frontier and matches models with 4x its active parameters on math, code, and long-context benchmarks.