Researchers release UEmbed multimodal embedding models
UEmbed produces sparse and dense multimodal representations in one decoder-only forward pass.
Why it matters
The work targets a core retrieval problem for search and RAG systems: combining lexical matching and semantic multimodal retrieval without separate cross-modal modules. If validated beyond the reported benchmarks, unified sparse-dense embeddings could simplify retrieval stacks for multimodal applications.
The key points
- 1.UEmbed supports sparse and dense embeddings in one model.
- 2.Models are released at 2B, 4B and 9B scales.
- 3.UEmbed-9B reports 71.8 dense and 71.0 sparse on MMEB-v2.
Researchers introduced UEmbed, a decoder-only multimodal embedding model that generates both sparse lexical and dense representations in a single causal forward pass. The team released 2B, 4B and 9B versions trained on public data. UEmbed-9B reports scores of 71.8 for dense retrieval and 71.0 for sparse retrieval on MMEB-v2, and remains competitive on BEIR.
⚡ Try this today
Evaluate UEmbed against your current dense and sparse retrievers before redesigning a multimodal RAG pipeline.
Sources & original reporting
This brief summarizes and links to reporting from the publishers below.
Enjoyed this brief? Get the next one in your inbox.
More in Research
Researchers introduce TTP-D for truck-drone routing
The paper models load-dependent collection routing with drone launches and rendezvous timing.
Study finds audit-repair context makes LLM verifiers more lenient
An arXiv paper reports lower false alarms after prior audit-repair episodes in model context.
AlphaEvolve helps lower matrix multiplication bound
A new arXiv note reports an improved upper bound for the matrix multiplication exponent.