Kolibri 1 drops sovereign 1M-token reasoning model

Kolibri 1 makes a credible case for sovereign, efficient AI deployments, though its hardware requirements limit local experimentation.

• –MoE sparsity keeps per-token compute closer to a small model while retaining larger-model capacity
• –Native German-English training and a tailored tokenizer target regulated European workloads
• –Long-context support, structured outputs, and tool calling make it practical for document-heavy agent systems
• –Apache 2.0 weights enable deployment flexibility, but the roughly 78GB FP8 footprint still demands serious hardware
• –Human-reviewed workflows remain the sensible use case; benchmark claims should be independently validated