A request to take a workload off a frontier API and run a downloadable model on the company's own servers usually arrives costed and unlicensed [11]. It reads like an infrastructure decision, and it is signed by someone who will later be asked to put in writing what the model was trained on.
That is the part the download does not cover. Weights arrive as weights, and the corpus behind them generally does not come with them [13], which turns provenance from something a customer can inspect into something a customer repeats on the supplier's authority. The freedoms a licence reviewer actually cares about, the ones about modifying an artefact and passing it on, belong to the third category's definition [14], and a model can sit comfortably in the second category while offering none of them. "Open weight" names how the thing reaches you, and stops there.
The vocabulary gap is visible if you count the examples. In the text as supplied to us, the proprietary section names three model families, the open weight section names four models, and the open source section reaches its definition without naming any [15]. Seven named products across three labels, and the label that carries verification rights has none attached to it.
For procurement, two questions do more work than the three labels. Are the weights in your custody? Can you evidence rights and provenance without quoting the supplier? That is a four-cell grid, and the taxonomy supplies names for three of the cells [16]. Cell one is no custody with supplier attestation, the plug-the-workflow-into-an-API arrangement where the contract is your audit trail [6]. Cell two is custody without provenance, the self-hosted open weight deployment where contracts and business plans stay on your hardware [11] while the training data stays unknowable [13]. Cell three is custody and provenance together, the shape the Open Source Initiative definition describes [14]. Cell four has no name in the taxonomy and is the easiest one to land in by accident: a downloadable model consumed through somebody else's endpoint, where you hold neither the weights nor the paperwork.
Here is what teams tell themselves when they pick cell two: open means we control it. Here is what the deployment does: it moves the unanswerable question from where the data sits to where the model came from, and the second question is the one an auditor or an enterprise customer asks in writing.
The forcing function is three sentences you have to be able to finish without the phrase "the supplier says". Where these weights came from and who published them. What this licence permits for a fine-tune we ship to customers. What we can show about the training data when a customer's security review asks. If a sentence only completes with an attestation, the model behaves like cell one no matter what the download page called it, and it should be papered and priced as a supplier dependency rather than booked as something you own.
