Kolibri: Europe's Open-Weight AI You Can Run Yourself
On October 3rd, German AI company Aleph Alpha released Kolibri, a 78-billion-parameter open-weight model that lets you run a capable AI entirely on your own infrastructure, without sending data anywhere. For creators and small companies in Europe dealing with the EU AI Act—which went live in August—that matters more than it sounds.
What's different here
Kolibri is a Mixture-of-Experts model with 78B total and 3B active parameters, supporting context up to 1M tokens, with full weights available on Hugging Face under Apache 2.0. The key difference: the entire model was developed and trained on infrastructure in Germany and Finland, adhering to European and German regulatory frameworks, including the EU AI Act and GDPR compliance by design.
A sovereign AI model is one an organization can run entirely on infrastructure it controls, built under the laws it has to follow, so its data stays in its hands and nobody else can change or switch off the model. That's the appeal: no API calls leaving your borders, no data handed to a US cloud provider, no terms-of-service negotiations.
Who actually wants this
Kolibri is built for sovereign mission-critical work in regulated areas including public administration, industrials and aerospace. But the open weights mean anyone can download it. It operates as a Mixture-of-Experts Transformer activating 3.46 billion parameters per token, with 40 layers using a 512-token sliding window and every fifth layer processing full context, enabling native support for up to one million tokens.
Its custom UniBPE tokenizer respects German morphological structures, reducing token count by 11 to 15 percent compared to standard tokenizers, which combined with sliding-window attention significantly lowers compute overhead for long-context retrieval-augmented generation.
The speed and the catch
The model beats rivals from spring, but current open-weight leaders are far ahead. That's honest: Kolibri isn't the fastest or most capable open model available today, but that's not the point. Kolibri landed as the top story on Hacker News on October 4 with 498 points and 298 comments—unusually high for a European AI company.
The practical constraint: the model card estimates about 78 GB of FP8 weight memory and specifies a minimum of two A100 80GB GPUs. Running it locally requires serious hardware. But the full weights can be downloaded from Aleph-Alpha/Kolibri-1 on Hugging Face, with FP8 and BF16 variants.
How to use it
Deployment uses Aleph Alpha's inference package and a Kolibri-specific vLLM plugin. If you have two A100s or better and care about keeping your work inside Europe: download the weights, spin it up locally. The model supports four reasoning settings (none, low, medium, high) plus tool calling, and was specialized for German, math, coding, long-context work, and agentic tasks.
For most creators, this won't be the everyday model—cost and complexity matter. But if you're building something that needs to stay compliant with European law, or you work in a regulated industry and need an AI that doesn't phone home, Kolibri is the first real, open option.