🤖Aleph Alpha Unveils Kolibri: 78.1B-Param LLM for German and English
A 78.1B-param LLM built in Germany, trained on 24 trillion tokens
TL;DR
Aleph Alpha's Kolibri, a 78.1B-param LLM, is trained on 24 trillion tokens, with over a fifth in German. It's designed for German and English, offering unique features like a sovereign status under European law.
Aleph Alpha just dropped Kolibri, a 78.1 billion parameter LLM for German and English. It's trained on 24 trillion tokens, with over 20% being German. If you're working on multilingual projects in Germany, this is a big deal. Kolibri's sovereign status means it's built and trained in Europe, under European and German law, making it a safe bet for teams wary of data sovereignty issues. It has 78.1 billion parameters, trained on 768 NVIDIA B200 GPUs, and its tokenizer needs 11.2% fewer tokens for German text than GPT-5's.
Key Points
Kolibri has 78.1 billion parameters, with 3.46 billion used per token.
Trained on 24 trillion tokens, with over 20% being German text.
Released on October 3, 2026, under the Apache 2.0 license.
Trained on 768 NVIDIA B200 GPUs, in Germany and Finland.
Tokenizer needs 11.2% fewer tokens for German text than GPT-5's.
Why It Matters
If you're working on multilingual projects in Germany, Kolibri's 78.1 billion parameters and sovereign status under European law make it a safer choice. Its tokenizer is 11.2% more efficient for German text than GPT-5's, but the real kicker is its data sovereignty. Teams with strict data residency requirements should take note.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,564 builders reading daily.