Kolibri: Aleph Alpha releases an open, 'sovereign' LLM that speaks German

On 3 October 2026, German company Aleph Alpha released Kolibri, a large language model built for German and English. Its weights are on Hugging Face under the Apache 2.0 licence. The details below come from an analysis published on the tej.as blog, which surfaced on Hacker News. That analysis draws on Aleph Alpha’s 189-page technical report, the model card and the launch announcement. Kolibri is German for hummingbird, which is a fitting name for a model whose whole selling point is being light. On paper, at least.

Heavyweight frame, featherweight footwork

Kolibri is a mixture of experts (MoE). In a classic “dense” model, every token passes through every parameter. In an MoE, each layer holds a crowd of small sub-networks, the experts, and a router decides which ones handle each token. Kolibri has 50 layers, each with 384 experts plus one shared expert that every token goes through. The router sends each token to 6 of those 384 experts. The upshot: 78.1 billion parameters in total, but only 3.46 billion actually doing any work per token, or 4.4%. Think of a 384-person office where six people answer each email and everyone else is in a meeting.

The rest of the spec sheet holds up well. The context window is 262,144 tokens natively, and it has been tested up to 1,048,576. The model offers four reasoning levels (none, low, medium, high), can call tools, and its knowledge stops on 18 June 2026.

There’s a catch, and the model card admits it upfront: the whole model has to sit in memory, even though only a fraction of it is working at any given moment. You’ll need roughly 78 GB of weights in FP8. So Kolibri computes like a 3.5-billion-parameter model but eats memory like a 78-billion-parameter one. The hummingbird may flap its wings at a dizzying pace, but it still needs a nest with room for all those experts.

What “sovereign” means here

In its announcement, Aleph Alpha gives the word two meanings. The first is about how it was made: the model was designed in Germany and trained on German and Finnish infrastructure, under European and German law, with “no foreign control”. The second is about how you use it: a government ministry or a car-parts maker can run Kolibri on its own servers. Its data never leaves the building, and nobody can tweak or switch off the model remotely. Aleph Alpha has signed the EU code of practice for general-purpose AI (GPAI) and says it designed the model with the AI Act in mind from day one.

Training covered about 24 trillion tokens, more than a fifth of them in German, on 768 NVIDIA B200 GPUs. The model card also concedes that not everything is homegrown. English text was rephrased using Google’s Gemma 4, German text with Mistral-NeMo, and Qwen3-32B labelled the data used for the quality filters. Aleph Alpha then filtered out the political biases those models can introduce, biases it says it measured on Chinese open models. Another caveat: only the weights and configuration files are under Apache 2.0. The training code and methods remain Aleph Alpha’s property. You can serve the dish to whoever you like, but you don’t get the recipe.

And how does it stack up against the Americans?

On that front, the available sources have little to offer. They contain no head-to-head comparison with the big US models. The only published result comes from Aleph Alpha’s own evaluation, in which Kolibri beats every model of its size it was compared against, in both German and English. Until someone independent runs the numbers, that performance is a claim made by the one party with the most to gain from it.

In any case, Kolibri’s value for European AI isn’t about topping a leaderboard. Its pitch is a model that’s strong in German, deployable in-house and built around European regulation, which the company sums up as “compliance comes as an inherited property”. For a public body that simply can’t ship its data to a foreign server, that often matters more than a few benchmark points.

So, for now, Europe’s AI sovereignty weighs in at 78 GB of FP8. Featherlight for a continent. Frankly obese for a hummingbird.