Aleph Alpha

Sovereignty, split in two

Aleph Alpha released Kolibri on 3 October 2026 and describes it as a sovereign open-weight model ^[1]. Open-weight means the parameters can be downloaded and run locally rather than reached only through a vendor's interface; sovereignty is the further claim the company makes about what that access is for.

The company divides sovereignty into two parts: how a model was built, and how it transfers to the customer ^[1]. The first half is about provenance. Aleph Alpha says it can account for every step from data ingestion through post-training to evaluation ^[1]. In practice that means a customer can ask where the training data came from, what was done to the model afterwards and how it was measured, and get an answer for each stage rather than a single final score.

The second half is about handover. A model that can only be reached on someone else's servers leaves the customer dependent on that operator, whatever the licence says. This consideration shapes most of the design choices below.

What Kolibri-1 is

Kolibri-1 has 78.1 billion total parameters and activates 3.46 billion of them per token ^[2]. The gap between the two figures is the point. Total parameters describe everything the model stores; active parameters describe what actually takes part in the computation for each token generated. A model of this shape can hold a large amount of learned material while paying the compute cost of a much smaller one at every step, which is what makes local deployment plausible on hardware a customer already has.

The language scope is equally deliberate. The model supports German and English only, a restriction the model card calls a choice of depth over breadth ^[2]: instead of spreading capacity across many languages, the training effort is concentrated in two.

The weights are released under Apache 2.0 ^[2], a permissive open-source licence that allows use, modification and redistribution, including commercially. The model card also states that Aleph Alpha signed the EU GPAI Code of Practice ^[2], the European Union's code for providers of general-purpose AI.

Built to run on the customer's own machines

Aleph Alpha pitches Kolibri's small active parameter count as what lets customers run it on-premise without sending data to a third party ^[1]. That closes the loop on the two halves of the sovereignty argument: a model whose construction can be traced, running inside the customer's own environment, with no data leaving the building.

The company positions Kolibri for mission-critical work in public administration, industrials and aerospace ^[1]. Settings like these tend to combine data that is expensive or impossible to move with a need for a model dependable enough to build on, which is the combination the on-premise argument addresses.

From Origin to Kolibri

Kolibri was not the first version of itself. It was first validated as Kolibri Origin, a 30 billion parameter model with 3 billion active parameters and a 65K context, before being scaled up ^[1]. Proving a training recipe at small size and then scaling it is a common way of reducing the risk of an expensive run, and naming the earlier stage puts that intermediate result on the public record instead of discarding it.

Where the bet lies

Aleph Alpha presents specialisation for German, rather than raw scale, as the main source of the model's advantage ^[1]. That is a different wager from the one implied by chasing benchmark leadership: the claim is that a model which knows one language and one regulatory environment well is more useful to the organisations working in it than a larger general model that knows many languages a little.

The trade is visible in the specification. A customer gets weights it can inspect, run and modify, an account of how the model was produced, and the ability to keep data in-house; in exchange the model covers two languages, and its size is framed around efficiency rather than around being the largest. For Aleph Alpha, sovereignty is not a feature bolted onto the model. It is the product being sold.