Aleph Alpha Kolibri

Released October 3, 2026, Kolibri is Aleph Alpha's 78.1-billion-parameter open-weights model for German-English document analysis, reasoning, and tool use, published under Apache 2.0. It is a mixture-of-experts design that activates only about 3.46B parameters per token, cutting compute per step while keeping the full weights downloadable for self-hosting. The headline spec is a 1,048,576-token maximum context, though the model card recommends staying at or below the 262K native training length for efficient serving and complex tasks. The published 8-bit weights run about 78GB, needing roughly two 80GB A100s or one H200/B200/B300 minimum.

Limitations: Very new (October 2026) with no independent evaluations yet — all benchmark and efficiency claims are vendor-supplied. German-English bilingual focus means other languages are secondary. Open weights do not eliminate the substantial hardware and operating costs of serving a 78B model, and Apache 2.0 alone does not certify EU AI Act compliance for any specific deployment.

Vendor-reported only; no independent evals as of October 2026