Aleph Alpha releases Kolibri, an open-weight model built to run on your own servers
German AI lab Aleph Alpha released Kolibri on October 3, an English-German mixture-of-experts model with 78 billion total parameters, of which only about 3.5 billion are active for each word it processes. It handles up to about a million tokens of context, supports tool calling, agent workflows and document Q&A, and is trained to say so when the supplied documents don't hold the answer. The full weights are free on Hugging Face under the Apache 2.0 licence, and it's pitched at industry and public administration that want to run AI on infrastructure they control. Aleph Alpha recently agreed to combine with Toronto's Cohere, pending regulatory approval.
What it means for manufacturers and fabricators
Capable open models you can run in-house keep getting cheaper to serve, so keeping your drawings, quotes and customer data off someone else's cloud is becoming a realistic option, not just a big-company luxury.