Kolibri: Germany’s open-source mixture-of-experts language model
Kolibri uses a custom MoE Transformer and is reported to outscore three competing models across several math and science benchmarks, though independent validation is still needed.
Germany has entered the frontier language-model race with Kolibri, an open-source model built on a custom mixture-of-experts (MoE) Transformer. This architecture activates selected parts of the network for each input rather than using the full model every time.
In the published comparison table, Kolibri scores above Qwen3.6 35B-A3B, Nemotron 3 Super and Mistral Small 4 across all six visible evaluations: the English and German versions of AIME 2025, AIME 2026 and GPQA Diamond. The reported scores include 96.9 on AIME 2025 English and 84.3 on GPQA Diamond English.
The image does not disclose the evaluation procedure, test conditions, precise system versions or independent validation. These figures should therefore be treated as results reported by the model’s promoters, not as conclusive evidence that Kolibri is superior across general use cases.