Aleph Alpha has come out with a new model called Kolibri.
What's very interesting about this model is that it was designed as a bilingual German-English model. According to Aleph Alpha, it was trained to reason directly in German rather than relying primarily on translated English material.
This could be of great value for German speaking companies, especially when working with German technical documentation, industrial procedures and compliance materials. Anybody who understands German and the ability of this language to describe complete processes with one word knows what I'm talking about. 😄
But I think there's a much wider potential here, particularly for organisations across the EU and even elsewhere. Kolibri also performs competitively in English, and a lot of emphasis was placed on data sovereignty, European regulatory requirements and documented training data governance.
Because it's an open weight model, organisations can deploy it on their own infrastructure. That could make it very interesting for companies dealing with sensitive information that cannot be sent to external cloud models, whether they're based in Germany, Ireland, UK, or somewhere else entirely. Of course, running a model locally doesn't automatically make the whole system secure or compliant, but having control over where the data is processed is a good starting point.
Another point worth mentioning is that Kolibri was trained for document grounded answering and abstention, meaning recognising when the supplied evidence isn't sufficient rather than simply inventing an answer. Aleph Alpha also reports improvements in this area, although I'd be interested to see how well that holds up in real-world applications. Anyway, you can tell its German roots. 🤣
The model does require some pretty beefy hardware, but for organisations with strict data requirements, that investment might be worthwhile, especially if they already have the infrastructure.
For more information, including the technical report, check it out here:
For more information you can check it out here: