Why a parameter count is not a hardware requirement
Two confusions account for most of the disappointment around open weights. One is about memory, the other is about rights, and both come from reading a model card as if it were a specification sheet. As of 2026-09-12.
A parameter count is a property of the model, and the memory a run consumes depends on precision, batch size, resolution, clip length and the surrounding pipeline. The same weights can fit comfortably on one card in one configuration and exhaust a much larger one in another, which is why the register records what the repository states rather than deriving a number from arithmetic that looks convincing.
The licence confusion is more expensive. A permissive licence on a set of weights governs the files: copying them, modifying them, redistributing them. Whether the video produced by those weights can be sold, and under what conditions, is addressed in a separate clause when it is addressed at all. Teams read the first and assume the second.
Both explainers exist because the mistakes are made by careful people. A model card is written by researchers for researchers, and a production team reads it looking for answers it was never meant to contain.
Where an entry in the register turns on either point, it links here rather than repeating the reasoning in a sentence that would have to be shorter than the argument needs.
1The explainers
- Reading a licence — A permissive licence on published weights answers fewer questions than it appears to.
- Planning hardware — Repositories state hardware in their own terms, and two figures rarely describe one run.
Neither page adds a family to the register; they set out how the entries should be read.
Also in the register: Models, Hosted, Questions, Data. What a release has to publish is set out on the reading page.