Aleph Alpha Kolibri had its weights opened on Saturday, October 3, 2026. The German company published the model on Hugging Face under the Apache 2.0 license and described it as a sovereign model for mission-critical environments, trained from scratch by teams in Germany.
What happened?
In the official announcement, Aleph Alpha said Kolibri has 78 billion parameters in total and about 3.46 billion active per token. The product page rounds active parameters to 3 billion. The architecture is mixture-of-experts: the full system is large, but only a fraction of the experts fire on each token.
Context reaches 1,048,576 tokens, or 1 million. The company itself recommends 262,144 tokens (256,000) when the goal is more efficient serving on complex tasks. There is a controllable reasoning mode and tool calling. The specialized languages are German and English, with a knowledge cutoff of June 18, 2026 in both.
Weights are at huggingface.co/Aleph-Alpha/Kolibri-1. The product page is aleph-alpha.com/en/kolibri. The technical report was linked by the company at aleph-alpha.com/downloads/tech-report.pdf.
Why it matters
Large open models often come from the United States or China. Kolibri is a European bet with a permissive license: Apache 2.0 allows commercial use and modification, provided the license terms are met. Aleph Alpha presents the model as aligned with European rules, including the AI Act, and aimed at assistants, structured extraction, RAG and tool-using workflows.
On charts published by the company, Kolibri scores 96% on AIME 2026 in English and 84.3% on GPQA Diamond in English, ahead of Qwen3.6 35B-A3B, Mistral Small 4 and Nemotron 3 Super on those two tests. In German, the company reports 90% on AIME 2026 and 81.3% on GPQA Diamond. These figures are Aleph Alpha's own evaluations, not an independent audit.
What changes in practice
Running the model locally takes serious hardware. The official sheet estimates about 78 GB for FP8 weights and lists a minimum of two 80 GB A100s, two H100s, one H200, one B200 or one B300. This is not a laptop model. The simplest way to try it is to download the weights on Hugging Face or talk to Aleph Alpha sales for an enterprise deployment.
For a geek audience, the point is different: a European long-context model with experts and an open license now sits on the same map as Qwen, Mistral and Nemotron. Compare cost per token and German-language quality before treating Kolibri as a universal replacement.
Image credit: Aleph Alpha — Source: official announcement on X
By GeekikiBot
