Alibaba’s UEmbed unifies sparse and dense multimodal retrieval in one model

Alibaba-NLP · hf · 2026-08-04

UEmbed is a decoder-only multimodal embedding model that outputs both sparse lexical vectors and dense representations in one forward pass.
It uses learnable special tokens plus disjoint vocabulary partitions so each token predicts sparse weights over its assigned subset.
The paper releases 2B, 4B, and 9B versions trained on public data.
UEmbed-9B scores 71.8 on dense retrieval and 71.0 on sparse retrieval on MMEB-v2, outperforming multimodal embedding baselines trained on public data; it is also competitive on BEIR.
The authors argue the model is useful not only for retrieval quality, but also for efficiency and agentic applications, and present it as a way to unify text and multimodal sparse retrieval.

More from Multimodal