KURE-sparse-nano

0.1B sparse retrieval model SPLADE outperforms dense embedding models in local RAG

A developer's tests found the sparse retrieval model SPLADE beat Qwen3-Embedding-0.6B on local RAG retrieval while using under 1.5GB of VRAM, running stably on a 16GB MacBook Air; the dense model overheated and rebooted the same machine. The author used the English-Korean KURE-sparse-nano and found that hybrid SPLADE-plus-dense retrieval performed worse than SPLADE alone.