German AI developer Aleph Alpha has released Kolibri, a German-English model whose full weights can be downloaded under the Apache 2.0 license. The company describes it as a mixture-of-experts model: it stores 78.1 billion parameters but activates about 3.46 billion for each token.
A mixture-of-experts model routes each token—a small piece of text processed by the model—through a selection of specialized neural-network blocks instead of calculating with every parameter each time. That can reduce per-token computation compared with using the entire model at once. It does not shrink the checkpoint to 3.46 billion parameters, however; a deployment still needs access to the full set of weights.
Aleph Alpha provides an inference package that adds Kolibri support to vLLM, an engine for running language models. The model card says the model can process up to 1,048,576 tokens in one context, but recommends no more than 262,144 for efficient serving and complex tasks. Reaching the one-million-token ceiling requires an explicit configuration change.
Self-hosting still depends on enough accelerator memory for the full checkpoint and on the context length an operator chooses; the active-parameter count alone does not establish a measured serving-cost reduction.
Aleph Alpha’s published benchmark comparisons are its own evaluations. Results vary by task and do not provide independent evidence of a general performance lead.
Sources
- Aleph Alpha: Kolibri release announcement
- Kolibri model card: architecture, serving instructions, context limits and evaluations
- Image: Carl Lender, “Datacenter Server Racks,” via Wikimedia Commons, CC BY 2.0. Representative data-center image; it does not show Aleph Alpha hardware.
.jpg){kind=link}