Aleph Alpha releases Kolibri, an open-weight German-English AI model

Aleph Alpha says Kolibri came out on 3 October 2026 and is an English-German mixture-of-experts model.[1][2] It has 78.1 billion total parameters, with 3.46 billion active per token, and supports context lengths of up to 1 million tokens.[2][3] The company made the full model weights available on Hugging Face under the Apache 2.0 license.[3] That makes the weights available for teams to examine and use under that license; the evidence does not specify hosted access, pricing or service-level terms.[3]

Aleph Alpha says its teams trained Kolibri end-to-end from scratch in Germany, using infrastructure in Germany and Finland.[4][1] It reports training the model on 24 trillion tokens, with German making up more than 20% of the data mix, and adding more than 2 trillion curated or synthetically generated German tokens.[2] The company says it specialized the model for German, reasoning, math and agentic behavior.[3] Its post-training reinforcement learning used more than 1.2 million internally curated tasks covering reasoning, tool use, instruction following, code and retrieval.[2]

Aleph Alpha says Kolibri is designed for regulated work in public administration, industry and aerospace, and intended to run on infrastructure customers control themselves.[3][5] The company says customers get deployment freedom and intellectual-property safety.[1] It also says it controls the development process across data, architecture, training infrastructure, post-training and evaluation.[2] These are company statements about the model’s intended use and development; the evidence does not establish how Kolibri performs in any particular customer environment or what implementation requirements customers would face.

Aleph Alpha says Kolibri sits on the quality-and-serving-cost Pareto frontier among open models it evaluated in English and German.[2] The company also reports that Kolibri scored above every compared model of its size in both languages in its own evaluations.[1] Aleph Alpha reported a 96.9% score on AIME 2025.[6] The evidence identifies these as company-reported results and does not provide independent validation or details sufficient to compare evaluation conditions across models.[2][1][6] Teams considering the model should therefore treat the claims as a starting point for their own testing.