UEmbed Delivers Unified Sparse and Dense Multimodal Embeddings
August 2, 2026
UEmbed is a decoder-only model that generates both sparse lexical and dense representations in a single causal forward pass. It uses learnable special tokens to predict sparse weights over partitioned vocabulary subsets, streamlining multimodal retrieval.
HOW THIS AFFECTS YOU
●
builderYou can reduce infrastructure complexity by using a single model for both dense and sparse retrieval.
●
researcherThis offers a new architecture for unifying multimodal embedding types within a causal framework.