Aleph Alpha Releases Kolibri: 78B MoE Model
Original: Kolibri: A Sovereign Open-Weight Model
Why This Matters
Open-weight sovereign MoE models targeting regulated European sectors are still rare.
German AI firm Aleph Alpha launched Kolibri on October 3, 2026 — a 78B-parameter Mixture-of-Experts model with 3B active parameters and 1M-token context, available on Hugging Face under Apache 2.0.
Aleph Alpha released Kolibri, an English-German Mixture-of-Experts Transformer with 78B total parameters and 3B active, supporting context windows up to 1 million tokens. The model is freely available on Hugging Face under the Apache 2.0 license. Kolibri follows Kolibri Origin, an earlier 30B-total/3B-active model with a 65k-token context window, which served as the pipeline validation stage. The company says the same end-to-end training pipeline — covering data ingestion and curation, ablations, pre-training, and post-training — handled both releases, enabling hundreds of ablation experiments and hardware-fault-tolerant stable runs without manual intervention. Kolibri is explicitly targeted at regulated sectors: public administration, industrials, and aerospace. Key specializations include German-language performance, reasoning, math, and agentic behavior. Aleph Alpha frames the release around what it calls 'sovereignty' — full supply-chain transparency from data sourcing through final evaluations, plus on-premise deployment freedom and IP safety for customers, which it says makes regulatory compliance an inherited property of the model rather than an afterthought. Its compact active-parameter count (3B) is pitched as enabling efficient on-premise inference without routing sensitive data to third-party services.