How would either an architecture or a model be copyrightable?

Architectures are algorithms. If they aren't patentable and are in addition to that patented, they have no protection.

Model weights are a result of a mechanical procedure that fits a model to data, minimising some kind of error. That is not a work of human authorship.

Things that could be copyrightable are an article describing a model architecture, or a specific software implementation of a model.

As an argument why model weights are unlikely to be copyrightable consider the following parallel: we know that model output, for example, a story generated by ChatGTP based on a prompt is certainly not copyrightable, since it's not a work of human authorship, but then, how is the model? We can view the selection of training examples as something similar to a prompt and the training process as similar to the inference. I think giving copyright protection to model weights might be reasonable though, but I think it's unlikely that they have copyright protection.