Perplexity releases pplx-embed-v2-late: OCR-free late-interaction embeddings topping retrieval benchmarks

perplexity_ai · x · 2026-10-08

Perplexity has open-sourced pplx-embed-v2-late on Hugging Face: two late-interaction embedding models that retrieve text, images, and rendered pages in a shared embedding space.

How it works

Unlike dense embeddings that compress a document into one vector, each token keeps a 128-dim vector scored with MaxSim, so query tokens match their closest document tokens.
Images and rendered pages are embedded directly, so PDFs, slides, and scans are searchable without OCR, preserving tables, figures, and layout.

Benchmarks

92.4% on MADQA (500 questions over 800 PDFs), the top retriever score.
On Q2D-Web (70k production queries over 190M pages), the 9B and 0.6B models hit 74.8% and 73.6% Recall@1000 vs 69.3% for nemotron-embed-8b.
Agentic search: a GPT-OSS-120B agent with the 9B model answers 64.0% of BrowseComp+ questions, 4.9 points ahead of ColBERT, with fewer searches than any baseline.
Both sizes are distilled token-by-token from one 18B teacher and share one embedding space: a corpus indexed with 9B can be queried with 0.6B, lifting ViDoRe v3 from 62.3% to 63.5% at no added query cost.

Original post →

Perplexity releases pplx-embed-v2-late: OCR-free late-interaction embeddings topping retrieval benchmarks

More from Models