Perplexity pplx-embed-v2-late: MIT embeddings for your PDFs

Perplexity released pplx-embed-v2-late, two open embedding models that search text, images and rendered document pages in one shared index. Both ship on Hugging Face under the MIT license, which means you can run them commercially on your own hardware. If your business sits on a pile of PDFs, scans and slide decks that nobody can search, this is the most practical embedding release of the month. It also comes with a storage bill you should price before you commit.

What actually happened

Perplexity announced the release on October 7. From the model cards for the 0.6B model and the 9B model:

• Two sizes, one space. A small model for cheap, fast queries and a large one for building high-quality indexes. They share an embedding space, so the small model can query an index the large model built.
• Late interaction. Each model outputs one 128-dimension vector per token and scores matches with MaxSim, ColBERT-style, instead of squeezing a whole document into one vector.
• Pages as images. The models embed images and visual documents directly. A scanned invoice or a chart-heavy deck is searchable without an OCR step.
• Benchmarks. On ViDoRe v3 (nDCG@10), the cards list 62.3% on images for the small model and 65.2% for the large one. These numbers are self-reported.
• No hosted API yet. The cards list no inference provider. For now, you host it yourself on a GPU.

Why it matters for your business

Most small-business "AI search" projects die in the OCR step. The text extraction from scanned forms, price sheets and supplier PDFs is bad, so retrieval is bad, so the chatbot makes things up. A model that reads the page image skips that failure point.

The MIT license matters too. You own the index and the model weights. No vendor can reprice your document search next quarter.

But read the trade-off before you index ten years of files. Late interaction stores one vector per token, not one per document. A long contract produces hundreds or thousands of vectors. Your vector database bill grows with document length, and not every database supports multi-vector search well.

Our approach for clients:

Pilot on one painful folder. The supplier catalogs or the scanned work orders. Measure whether staff find the right page faster.

Index with the big model, query with the small one. The shared space is the point. Build the index once on a rented GPU, then serve queries cheaply.

Price storage first. Multiply your page count by the average token count per page. That is your vector count.

Keep the source files. Embeddings are an index, not an archive.

Sources: Perplexity announcement on X, pplx-embed-v2-late-0.6b model card, pplx-embed-v2-late-9b model card.