On 3 October 2026, Aleph Alpha released Kolibri, a German-English model with openly available weights. The release offers a useful starting point for a business discussion that is often reduced to model rankings: how much control does an organisation need over an AI capability it expects to rely on?

What happened

Aleph Alpha’s model card lists approximately 78 billion total parameters and publishes the weights and configuration under Apache 2.0. It also describes hardware requirements and intended uses involving human review. Open weights provide a deployment option; they do not mean that the entire development process is open, that operation is free or that a particular business use is automatically compliant.

The distinction keeps the strategic question realistic. A company can gain choices about where and how to run a model while taking on work that a hosted service would otherwise perform. Control has an operating cost as well as an appeal.

Why it matters

Consider a hypothetical business with an internal document assistant that has become important to daily work. If its provider changes price, withdraws a feature or experiences a prolonged interruption, can the company keep the service running? A downloaded model is only one possible part of the answer.

The organisation also needs its approved documents, retrieval configuration, evaluation cases and application logic. Somebody must know how to operate the alternative and test whether it still gives acceptable answers. Otherwise, the apparent exit exists on a presentation slide but cannot be used under pressure.

This is why portability should be tested as a business capability. Take a contained task with non-sensitive test material and compare the current arrangement with a realistic alternative. Record what transfers easily, what has to be rebuilt and which trade-offs users notice. A slower but workable fallback may be valuable; an untested promise is harder to price.

The bigger shift

The decision between a hosted service and a model under the organisation’s control is not a contest with one universal winner. A hosted option may make sense where convenience, support and frequent improvements matter most. Greater deployment control may deserve consideration where continuity, location or customisation requirements justify the additional responsibility.

The comparison should include staff capacity, infrastructure, updates, security and the cost of evaluating changes. It should also distinguish a right to use the weights from ownership of every component around the service. A business can remain dependent on specialist expertise or another software layer even after changing the model provider.

AI strategy should make these dependencies explicit. The objective is not independence for its own sake. It is an arrangement in which the organisation understands what it is renting, what it controls and what would happen if an important assumption changed.

My take

Ask the team responsible for a critical AI workflow to demonstrate one modest exit test. Export the materials the organisation is entitled to retain, run a representative evaluation against an alternative and document the gaps. Keep the exercise small enough to complete, rather than turning it into a second transformation programme.

The result may support staying with the existing provider. That is still useful: the decision would rest on evidence about switching costs and service quality instead of an assumption that leaving is impossible or effortless.

Open weights expand the choices available to a business. Turning those choices into bargaining power or operational resilience takes preparation. The valuable asset is not merely the ability to download a model. It is the ability to keep an important process working when the terms around that model change.