The model is designed for sovereign, mission-critical work in government and regulated industries, with grounding features and sector-specific evaluations.
Kolibri was trained on 768 B200 GPUs, starting with 20 trillion tokens over 21 days, and reached nearly 24 trillion tokens total.
The model uses 384 experts with six active per token, full attention in 10 of 50 layers, and a 512-token sliding window in the other 40 layers.
German accounts for 21.3% of pre-training tokens, supported by a bilingual 128,000-entry vocabulary and a tokenizer designed to preserve German compounds.