Aleph Alpha releases Kolibri-1, a 78B sovereign open-weight MoE, on Hugging Face

Aleph Alpha released full weights of Kolibri-1 on Hugging Face under Apache 2.0. It is a mixture-of-experts with 78 billion total parameters but only 3.46 billion active per token, a context window of 1,048,576 tokens, and training on 20 trillion tokens with native float8 block quantization.

It supports explicit chain-of-thought and function calling, and is pitched as a sovereign European reasoning model optimized for German and English.

The tiny active-parameter count makes it cheap to run, fitting the week's sovereignty and local-inference themes alongside Mistral's Munich hub.

Skeptics say 'sovereign AI' framing masks benchmark losses to Qwen 3.8 27B. Watch independent evals and enterprise uptake in Germany.