Perplexity pplx-embed-v2-late
Open MIT-licensed late-interaction embeddings that search PDFs and slides without OCR.
Quick verdict
- Best for
- RAG over document scans, slides, and PDFs
- Pricing
- Free open weights (MIT); hosted API forthcoming
- Not ideal if
- Multi-vector indexes are larger and need a compatible vector store
What is Perplexity pplx-embed-v2-late?
On October 7, 2026 Perplexity released pplx-embed-v2-late, two open-weight late-interaction embedding models (0.6B and 9B parameters) under the MIT licence on Hugging Face. Built on Qwen3.5 with bidirectional attention, they emit one 128-dimensional vector per token and score with MaxSim, embed text, images, and rendered PDF pages, and share one embedding space so an index built with the 9B model can be served by the 0.6B one. A hosted API was announced as forthcoming.
Key features
- Late-interaction (multi-vector) retrieval with MaxSim
- Text, image, and PDF-page inputs
- Shared embedding space across 0.6B and 9B sizes
- MIT licence on Hugging Face
- Perplexity-reported ViDoRe v3 scores (62.3% / 65.2%)
Pros
- Searching slides and PDFs visually avoids an OCR pipeline
- Index once with the big model, serve with the small one
- Permissive licence
Cons
- Multi-vector indexes are larger and need a compatible vector store
- One input cannot mix text and images
- Benchmarks are vendor-reported
Best for
Read more
Alternatives to Perplexity pplx-embed-v2-late
LlamaIndex
The data framework for connecting LLMs to your documents, plus LlamaParse for hard PDFs.
Qdrant
A fast, Rust-built vector database with rich filtering and a generous free cloud tier.
Weaviate
Open-source vector database with hybrid search you can self-host or run in the cloud.