Multimodal embeddings beyond a single vector (17 minute read)

TLDR AI Models
embeddings multimodal retrieval perplexity vector-search open-weights

Summary

Perplexity 发布了 PPLX-EMBED-V2-LATE，提供多向量、多模态嵌入，保留 token 级向量并在共享嵌入空间中实现更丰富的文本与图像检索，在 ViDoRe(V3) 等基准上表现领先。

The Perplexity team launched PPLX-EMBED-V2-LATE, providing multi-vector, multimodal embeddings that enable richer retrieval across text and images. These models retain token-level vectors and utilize a shared embedding space, leading to improved retrieval on benchmarks like ViDoRe(V3). Available in two sizes, 0.6B and 9B, they offer industry-leading performance for various retrieval tasks while maintaining efficiency.
Original Article

Similar Articles

UEmbed: Unified Sparse and Dense Multimodal Embeddings

Hugging Face Daily Papers

UEmbed is a decoder-only multimodal embedding model that produces both sparse and dense representations in a single forward pass, released at 2B, 4B, and 9B scales. It outperforms existing public-data-trained multimodal embedding models on MMEB-v2 and remains competitive on BEIR.

WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report

Hugging Face Daily Papers

WeMM-Embedding is a family of universal multimodal embedding models developed by Tencent, achieving state-of-the-art performance in retrieval and recommendation tasks for WeChat applications, with variants in 2B, 4B, and 9B parameters released publicly.

Learning Multimodal Embeddings with Evidence-Aligned Readout

Hugging Face Daily Papers

这篇论文提出 EviAlign，将语义证据生成与边界读出（Boundary Readout）耦合到共享的多模态大语言模型中，联合训练生成与对比检索目标，在 12 个 MMEB 检索任务上以 500K 训练对达到 76.9 的平均 Recall@1，同时保持单向量索引与打分。