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?
The current status quo seems to be
[Model name]-[Major version]-date-[minor variant/feature flags]
E.g. [Claude Sonnet] [3.7] [2025-03-02]:thinking
with a few differences based on the specific company.
Aside from some silliness (e.g. OpenAI's naming schemes; Anthropic deciding to go with "New Sonnet 3.5" rather than "Sonnet 3.6"), I think it works reasonably well.