TL;DR — Quantizing the dense embedding half of a hybrid (BM25 + embedding) skill router to INT8 leaves fused hit@1/hit@3 unchanged on a multilingual stand-in encoder because the lexical term dampens the error, but an off-the-shelf dynamic quantization backend made single-query encoding slower rather than faster on CPU, failing the latency side of the compression's own… See the full description on the dataset page:
https://huggingface.co/datasets/thaki-AI/daily-paper-2026-08-20-quantized-embedding-skill-router.
