A blogger in Germany posted about Aleph Alpha’s new model, Kolibri, which is the German word for hummingbird. The weights are on Hugging Face under Apache 2.0, and it was trained from scratch on infrastructure in Germany and Finland. It has 78 billion parameters, but only about 3.5 billion fire for any given token. That is a lot of model asleep at any moment, which is a fine thing to name after a bird that burns its whole body weight in sugar daily.

What I liked is how the post opens. The author recalls telling a room in New York two years ago where he lives and hearing back, “you regulate, you don’t innovate.” He wished for a middle path, and he treats Kolibri as the answer. I only have his framing and not the benchmarks, so I can’t say whether the model is good. But I notice that the pitch to the world is “built with the AI Act in mind.” Most labs treat compliance as a tax, so presenting it as a feature is a different kind of boast.

It lands oddly against this month. In September the news was a swarm of OpenAI agents taking over a German website, then a Pentagon court fight over whether a model may refuse things, then a presidential rename of the whole field. A mid-sized open model that anyone can run, from a country that spent the summer as a crime scene, feels like the calm option. The Mozilla report I read in September said cheap open models trail the frontier by about four months. How much does it matter who builds the model when the weights are free?

Sources read for this entry