Details
- Perplexity released two late-interaction embedding models, pplx-embed-v2-late in 0.6B and 9B sizes, for retrieving text, images, and rendered pages.
- Both models are publicly available on Hugging Face and use a shared embedding space, enabling a corpus indexed with the 9B model to be searched using the smaller 0.6B model.
- Instead of compressing each document into one vector, the models retain a 128-dimensional vector for each token and use MaxSim to match query tokens with their closest document tokens.
- The models directly embed images and rendered pages, allowing PDFs, slides, scans, tables, figures, and page layouts to be searched without OCR.
- Both models were distilled token by token from a single 18B teacher model, preserving compatibility between the two sizes.
- On ViDoRe v3, searching a 9B index with the 0.6B query model improved performance from 62.3% to 63.5% without increasing query cost.
- On Q2D-Web, which covers about 70,000 production-derived queries across 190 million web documents, the 9B and 0.6B models achieved Recall@1000 scores of 74.8% and 73.6%, compared with 69.3% for nemotron-embed-8b.
- A GPT-OSS-120B agent using the 9B model answered 64.0% of BrowseComp+ questions correctly, while MADQA results reached 92.4% across 500 questions over 800 PDFs.
Impact
The release targets a practical retrieval bottleneck: large models improve indexing quality, while smaller models reduce query-time cost. Shared embeddings let organizations separate those workloads without rebuilding their index, and direct page and image retrieval reduces dependence on OCR for visually structured documents. The reported benchmarks position Perplexity among active competitors in web-scale and multimodal retrieval, although results vary by dataset and metric.
