The company says the model has 78 billion total parameters, about 3 billion active parameters per token, and supports context windows up to 1 million tokens.
Full Kolibri weights are downloadable through Hugging Face under Apache 2.0. Aleph Alpha says serving is available through its aleph-alpha-inference package, with a vLLM plugin or container image. The release follows Kolibri Origin, a 30B-total, 3B-active model with a 65k-token context window.
Aleph Alpha reports benchmark results including 96.9 on English AIME 2025, 87.5 on German AIME 2025, and 85.9 on LiveCodeBench v6. It also says Kolibri is designed for on-premise use in public administration, industrial, and aerospace settings. The announcement does not establish independent benchmark validation, real-world deployment results, inference cost, or hardware requirements.