Aleph Alpha's Kolibri: 78B-Parameter Sovereign EU LLM Ships Under Apache 2.0
Aleph Alpha shipped Kolibri 1 on October 3, the Heidelberg-based lab’s first open-weight flagship since going all-in on European sovereign AI. The model is a mixture-of-experts architecture with 78.1 billion total parameters but only 3.46 billion active per token — a 4.4% activation ratio that brings the per-token compute cost down to roughly a 3.5B dense model, while the full 78B parameter set stays resident in memory.
Specs
| Field | Value |
|---|---|
| Total parameters | 78.1B |
| Active per token | 3.46B (4.4%) |
| Context | 262,144 tokens native; tested to 1,048,576 |
| Languages | German and English |
| Reasoning | None / Low / Medium / High (4 levels) |
| Tool calling | Yes |
| Knowledge cutoff | June 18, 2026 |
| Memory (FP8) | ~78 GB |
| License | Apache 2.0 (weights and config) |
| Training compute | 768 NVIDIA B200 GPUs |
| Training data | ~24 trillion tokens; more than a fifth German |
The Sovereignty Pitch
The word “sovereign” carries specific meaning for Aleph Alpha. Training happened on infrastructure physically located in Germany and Finland, under European and German law, with no foreign-entity control over the process. Weights are available under Apache 2.0, which means any customer can run the model on-premise with data that never leaves their building.
That matters in regulated verticals. A German federal ministry or automotive supplier can deploy Kolibri, know exactly what laws govern it, and face zero risk of a US executive order or vendor policy change pulling the model from under them. Aleph Alpha has also signed the EU’s General-Purpose AI Code of Practice, giving enterprise procurement teams a compliance paper trail.
Aleph Alpha’s internal evals place Kolibri above every model of comparable size in both German and English — though that benchmark is self-reported and has not yet been independently reproduced. The lab published a 189-page technical report alongside the release.
Why This Matters
Open-weight models in Europe with serious German-language coverage have been thin. Most labs train with token budgets that reflect the Internet’s English dominance; German-language performance typically degrades proportionally. Kolibri’s >20% German training token share is a deliberate correction.
The B200-based training stack gives Aleph Alpha a current-generation compute story — NVIDIA’s Blackwell architecture offers significant memory bandwidth improvements over H100 that translate directly to MoE efficiency. At 78 GB in FP8 (8-bit floating point), Kolibri’s weights fit on a single B200 GPU (192 GB VRAM) or a single H200 (141 GB VRAM), keeping inference hardware requirements low relative to the model’s raw parameter count.
The four-level reasoning mode is operationally useful: enterprises can trade inference cost against task complexity, running at “none” for classification pipelines and “high” for complex multi-step reasoning, all from the same weight file.
Weights are on Hugging Face at Aleph-Alpha/Kolibri-1.