Aleph Alpha releases Kolibri-1 weights under Apache 2.0

Aleph Alpha has released the full weights for Kolibri-1, an English-German mixture-of-experts reasoning model aimed at sovereign and self-hosted deployment. The publisher lists 78.1 billion total parameters and 3.46 billion active parameters per token, with 384 experts and six selected per token. Kolibri is available in FP8 and BF16 variants, and Aleph Alpha states that the weights can be used under Apache 2.0 terms. The model card also lists explicit reasoning modes and tool calling, making this a direct new-model event rather than a hosted-API announcement.

Long context needs careful wording. Aleph Alpha states a context length of 1,048,576 tokens, but its release table says the longest trained length was 262,144 tokens; the model card recommends 262,144 tokens or less for serving efficiency and complex tasks. The architecture combines a 512-token sliding window with full attention every fifth layer. Aleph Alpha says pre-training used 20 trillion tokens and that the knowledge cutoff for both English and German is June 18, 2026. These publisher disclosures are useful registry evidence, but the trained-length and serving-context fields should not be collapsed into one inferred value.

Aleph Alpha also publishes benchmark results across math, coding, grounding, agentic and long-context tasks and says Kolibri sits on a quality-versus-serving-cost Pareto frontier for English and German. Those measurements were run with the company's own harnesses and should remain explicitly attributed rather than treated as independently reproduced rankings. For Open Model Weights, the stronger signal is the release package itself: downloadable weight variants, a stated permissive license, detailed architecture and training metadata, and deployment guidance. Kolibri therefore clears the bar for immediate registry review, with the FP8 and BF16 artifacts checked separately and missing fields left unknown.