Aleph Alpha’s Kolibri model addresses the core pain points of organizations that need powerful AI while staying inside strict regulatory and operational boundaries. Public administrations, industrial firms and aerospace contractors often struggle with models that are either too large to run on existing hardware, too opaque for compliance audits, or poorly tuned for German‑language documents. Kolibri solves these issues by delivering a 78.1‑billion‑parameter Mixture‑of‑Experts system that activates only 3.46 billion parameters per token—just 4.4 % of its total weight. This sparsity lets the model run on a single NVIDIA B200, B300 or H200 GPU, or on two H100 SXM5 cards, using the vLLM serving stack with dedicated reasoning and tool‑call parsers.
Because the checkpoint is released under the Apache 2.0 license, teams can integrate Kolibri into sovereign environments without worrying about restrictive licensing or vendor lock‑in. The model was trained entirely on infrastructure in Germany and Finland, with a data pipeline that redacts personal information before training, helping it meet EU AI Act requirements, the General‑Purpose AI Code of Practice, and GDPR.
Kolibri’s hybrid attention design combines 10 full‑attention layers with 40 sliding‑window layers that keep a fixed‑size KV cache, making a 1‑million‑token context feasible without exponential memory growth. Its German‑specific UniBPE tokenizer yields roughly 11 % fewer tokens on German text compared with generic alternatives, lowering latency and cost for local language workloads.
Users can adjust reasoning effort per request—from a no‑think mode to three graduated levels—so latency and compute spend match the complexity of each task. The model is ready for download on Hugging Face, enabling rapid deployment in on‑premises or private‑cloud settings.
In short, Kolibri gives regulated sectors a high‑performance, compliant, and cost‑effective AI solution that fits existing hardware, supports long German‑English contexts, and lets organizations control reasoning depth on a per‑request basis.
#AI #LanguageModel #MoE #GenerativeAI #RegTech #SovereignAI