V-SPLADE

"Inference-Free Multimodal Learned Sparse Retrieval for Production-Scale Visual Document Search" from NAVER, SNU, and Naver Labs Europe: a 0.25B sparse retriever for visual documents on a ModernVBERT backbone where queries are resolved by learned bag-of-words lookup with no neural query encoding at serving time (inverted index, no GPU). Documents are embedded directly from pixels, reported 20x+ faster than OCR pipelines. Two checkpoints: v-splade-quality (avg NDCG@5 60.1 across six visual-doc benchmarks, +13.8pp over same-scale dense baseline) and v-splade-efficient (56.4); more than 2x R@5 over a same-backbone dense retriever on an 18.7M-page production corpus. Apache-2.0.