Kolibri: A Sovereign Open-Weight Model
Points and comments are a snapshot, not live.
Aleph Alpha's Kolibri is a sovereign open-weight model for regulated German/English work.
Kolibri is a 78B total, 3B active parameter Mixture-of-Experts transformer supporting 1M tokens. Released on German Reunification Day under Apache 2.0. Built for mission-critical work in regulated areas (public administration, industrials, aerospace) with a focus on German (21.3% of pre-training tokens), reasoning, math, agentic behavior, and grounding (trained to say "I don't know"). Benchmarks show it matches models with up to 4x its active parameters. Training pipeline runs autonomously, handling hardware failures. Two dimensions of sovereignty: supply-chain integrity and customer deployment freedom.
What commenters are saying
Commenters are skeptical of the model's value beyond niche use cases. Some question what it excels at: one top commenter says it knows less from memory, multi-turn tool calling is weaker, and it's not the best coding agent. Others note its strength is reading German documents and being one of the least hallucinating models. A second camp argues sovereignty matters for long-term independence from private sector whims. Multiple jokes circulate about the model overthinking and being good at faxes, playing on stereotypes about German bureaucracy.