Aleph Alpha Kolibri: How the sovereign German LLM works

414 points · 12 comments on HN · read original →

Points and comments are a snapshot, not live.

Aleph Alpha released Kolibri, an open-weight German/English LLM with 78B total but 3.5B active parameters per token.

Kolibri is a mixture-of-experts LLM with 78.1 billion total parameters and ~3.5 billion active per token, trained on 24 trillion tokens on 768 NVIDIA B200 GPUs. It uses a custom tokenizer requiring 11-24% fewer tokens for German text than alternatives, supports up to 1 million tokens of context, and offers four reasoning effort levels. Aleph Alpha claims it scores above all compared models of its size in both German and English. The model is licensed under Apache 2.0, with weights on Hugging Face, and was trained on European infrastructure with a focus on EU AI Act compliance.

What commenters are saying

Commenters noted the post was mislabeled as Show HN since the author didn't create the model. One commenter from the training team highlighted the model's strong iteration velocity and noted more releases are coming. A question was raised about why new models don't adopt DeepSeek's KV tweaks for cheaper inference, asking about potential drawbacks.