Aleph Alpha releases Kolibri, a 78B German-English open-weight MoE under Apache 2.0

Why it matters

【Key points】

Aleph Alpha released Kolibri on October 3, 2026, as an English-German Mixture-of-Experts model with open weights on Hugging Face under Apache 2.0. The model card gives 78B total parameters and 3.46B active per token.

Context is 262,144 tokens natively and 1,048,576 by extrapolation. Aleph Alpha positions Kolibri for sovereign, regulated work and states that it is a signatory of the EU GPAI Code of Practice.

【Verifiable】

Primary: https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/ and the model card at https://huggingface.co/Aleph-Alpha/Kolibri-1 .

Per the model card: 20T pre-training tokens (~62.5% English, ~23.9% German, ~13.6% code); pre-training ran 21 days on 768 NVIDIA B200 GPUs at 6.4e23 FLOPS, on infrastructure in Germany and Finland. FP8 weights take ~78 GB of memory.

The blog's score table, against Kolibri Origin, Qwen3.6-35B-A3B, Nemotron 3 Super 120B-A12B and Mistral Small 4 119B-A6B: AIME 2025 96.9 and 87.5 in German; GPQA (diamond) 84.3 and 81.3 in German; LiveCodeBench v6 85.9. It says Kolibri "matches models with up to four times its active parameter count, such as Nemotron 3 Super".

【Our read】

Europe's sovereignty pitch in practice: open weights, a permissive licence, European compute and a signed code of practice, sold on compliance and deployability rather than frontier scores. The same week Mistral announced Large 4, with weights due at the end of the month.

The 3.46B active parameters are the real trade-off: compute per token is cheap, but all 78B must sit in memory, so the saving is arithmetic, not hardware. It suits private, regulated, long-document work more than frontier use.

German is the differentiator, but scores are self-reported, and the card's wider table is less flattering: on overall English score Kolibri's 75.5 trails the dense Qwen3.8 27B at 80.2.

The voluntary GPAI code does not by itself settle AI Act duties. Watch for independent evaluation and whether 1M context holds up.