Open-weight models cross the enterprise threshold

The gap between the best hosted models and the best open-weight models has narrowed to the point where, for the specific workloads finance functions run, it is frequently not observable. Extraction from invoices and contracts, classification, summarisation and structured output generation all sit comfortably within the capability of models an organisation can run itself.

The economics are less clear-cut than advocates suggest. Self-hosting replaces a variable per-token cost with a fixed infrastructure and engineering cost. Below a meaningful volume threshold the hosted API is cheaper all-in once you price the platform engineering time honestly, which organisations consistently fail to do. Above it, the saving compounds quickly.

The stronger argument for self-hosting is not price. It is that certain documents — draft accounts, board papers, transaction materials, personnel data — should not leave your infrastructure regardless of contractual assurance. That argument does not depend on distrust of any provider; it depends on the observation that a control you enforce technically is stronger than one you enforce contractually.

For most mid-market finance functions the practical answer remains hybrid: hosted models for general work, self-hosted for the confidential subset. What has changed is that the confidential subset no longer means accepting materially worse output, which was the honest objection eighteen months ago.