Kolibri 1

  **The sovereign European MoE, one year after it needed a state subsidy to exist.** Aleph Alpha shipped Kolibri 1 on October 3, 2026 (German Unity Day): 78B total parameters, 3.46B active per token, trained on a 20T-token bilingual corpus (62.5 percent English, 23.9 percent German, 13.6 percent code). Apache-2.0 weights on Hugging Face; the company is an EU GPAI Code of Practice signatory and markets the release as the first sovereign European open-weight MoE.

  **The architecture is the efficiency story.** 50 layers split 40 sliding-attention (513-token window) to 10 full-attention; positional encoding lives only in the sliding layers, so the context extends to 1,048,576 tokens without the KV cache exploding: the 10 full-attention layers run about 20.5MB per 1,000 tokens at fp16 and the 40 sliding layers cap at roughly 21MB fixed, so a 256k-token context holds about 5.3GB of cache. MoE 384 experts with 6 active plus one shared expert on a 2,560 hidden size.