A vendor pitches a multilingual retrieval encoder at 27...

A simple comparison of parameter counts is misleading. One must break down the architecture; for instance, if both models share the same hidden dimension and layer count, the difference in parameter count is often just an expanded embedding table. In such cases, the encoding throughput remains identical, and the decision rests solely on whether the multilingual coverage provided by the larger embedding table justifies the additional resident memory cost per replica.