Dongcha Beating AI News, German AI company Aleph Alpha released and open sourced the English-German bilingual model Kolibri. The model adopts MoE architecture with a total of 78.1 billion parameters, but each token only activates about 3.46 billion parameters. The weights have been uploaded to Hugging Face, using the Apache 2.0 protocol.
Kolibri supports tool invocation and tunable inference modes, with context scaling up to approximately 1 million tokens. However, the official recommendation is that daily deployment should be controlled within 262,000 tokens, and the longest context actually trained by the model is also 262,000.
In the official evaluation of Aleph Alpha, Kolibri scored 96.9% in AIME 2025 and LiveCodeBench v6 scored 85.9%. The comprehensive English and German scores are slightly higher than Qwen3.5 35B-A3B, which also has about 3B activation parameters, but still lower than Qwen3.8 which has 27B full activation.