Open-weight model

Definition

An AI model whose trained weights are made available so others can download and run it.

Put simply

Instead of only accessing the model through its creator's service, an organisation can potentially operate or adapt the model itself.

What boards and leaders should understand

Open-weight models can create greater control over deployment, data handling, customisation, model versions and vendor dependency.

But control comes with responsibility.

Where an organisation operates the model itself, it may take on responsibilities that a managed provider would otherwise carry, including infrastructure security, safeguards, monitoring, updates, maintenance, evaluation and incident response.

So the strategic question is not simply “Could we run our own model?” It is also “Do we want to become responsible for operating it well?”

Questions to consider

  • Where would greater control over the underlying model create enough strategic value to justify greater operational responsibility?