Kolibri: A Sovereign Open-Weight Model

Aleph Alpha released Kolibri, a 78B-total, 3B-active mixture-of-experts transformer for English and German, with 1M-token context and Apache 2.0 weights on Hugging Face. Two models shipped three months apart because the pipeline runs as code: every change triggers an end-to-end run, checkpoints land hourly, and 38 hardware faults in 21 days went unattended. Specialised for German and agentic work, it is trained to abstain, holding back on 44% of AA-Omniscience items.

The discussion

Commenters mostly distrust the benchmark framing: the comparison set is a year old, and spijdar and amoshebb note Qwen3.8 27B beats Kolibri on German in Kolibri’s own harness, 79.9 to 70.8. The sovereignty framing is challenged head-on — petesergeant calls it a cover for weak performance, and others ask how sovereign a Cohere-owned, Toronto-operated company is. The team and its defenders answer that sovereignty is choice and control of data, not benchmark position, and that small tool-calling models need less copyrighted data than people assume. A long sub-thread splits on whether open weights are neo-feudalism in reverse or a proliferation risk nobody can gate.