> we don't want to hurt performance on other real-world tasks just to do well on MTEB

Nice!

Fortunately MTEB lets you sort by model parameter size because using 7B parameter LLMs for embeddings is just... Yuck.