You always want a bit of headroom for context. It's a problem I keep bumping into with 32B models on a 24GB card: the decent quants fit, but the context you have available on the card isn't quite as much as I'd like.