Ask HN: How should we version our models?

AI companies have been notorious for inconsistent model versioning schemes, but this still begs the question: how should we be versioning our models?

For traditional software, we have SemVer which is relatively straightforward. But for models (AI, statistical and otherwise), it's not as simple. When it comes to modelling, it is possible to improve upon a model without a) making backward incompatible changes, b) adding new features or c) fixing a bug... so how should we go about naming our models when we've improved them?