The ultra-simplified napkin math is 1 GB (V)RAM per 1 billion parameters, at a 4-5 bit-per-weight quantization. This usually gives most of the performance of the full size model and leaves a little bit of room for context, although not necessarily the full supported size.
Yes, it's more a rule of thumb than napkin math I suppose. The difference allows space for the KV cache which scales with both model size and context length, plus other bits and bobs like multimodal encoders which aren't always counted into the nameplate model size.
