According to CoinMeta, Aleph Alpha has released and open-sourced a bilingual model in English and German named Kolibri. This model utilizes the MOE architecture and has a total of 78.1 billion parameters, but only about 3.46 billion parameters of each token are actually activated. The weights have been uploaded to Hugging Face, and it adopts the Apache 2.0 protocol. Kolibri supports tool invocation and adjustable inference modes, with the context capacity expandable up to approximately 1 million tokens. The official recommendation is to keep the daily deployment within 262,000 tokens. Based on the official evaluations by Aleph Alpha, Kolibri scored 96.9% in AIME 2025 and 85.9% in LiveCodeBench V6. Its overall performance in both English and German is slightly higher than that of Qwen3.5 and A3B, which also have around 30 billion activated parameters, but it is still lower than that of Qwen3.8 with 27 billion fully activated parameters.