Aleph Alpha has released Kolibri 1, a bilingual German-English model published under the Apache 2.0 license with full weights on Hugging Face. It is a mixture-of-experts transformer with 78B total and 3.46B active parameters, a native context window of 262,144 tokens that can be extended to 1,048,576, and a custom UniBPE tokenizer tuned to German word structure and compound nouns; German accounted for 21.3% of the 20T pre-training tokens. Aleph Alpha says the model was built with the EU AI Act, the General-Purpose AI Code of Practice and the GDPR in mind, and trained with its in-house Merlin-Arthur protocol so it abstains when the supplied context does not support an answer. On the company’s own benchmark table, Kolibri leads smaller open-weight peers on math and long context but trails the much smaller Qwen3.6-35B-A3B on closed-book knowledge (AA-Omniscience Index of -32.8 against -15.3) and on function calling (BFCL v4 overall 61.4 against 67.2).

Related: Mistral Releases Medium 3.5 and Remote Agents in Vibe Environment, Thinking Machines Lab Releases Inkling, Their First Open-Weights Multimodal Model

Kolibri has landed: a sovereign open-weight model