Perplexity pplx-embed-v2-late: what it is and is it free?

pplx-embed-v2-late is a pair of open-weight search models from Perplexity, a 9-billion-parameter model and a 0.6-billion one, built to find the right page, paragraph or image in a pile of documents. Both are free to download from Hugging Face under the MIT license, which allows commercial use. The catch is that Perplexity doesn't host them yet, so you have to run them yourself.

Perplexity announced the release on October 7 in a post from @perplexity_ai, calling it "two late-interaction embedding models" that retrieve text, images and pages in one shared space. Within the hour, Perplexity search engineer @denis_bykov posted "Please welcome pplx-embed-v2-late! Extra strong late interaction models," with a pointer to Hugging Face and credit to the team that built it.

It's Perplexity's second open release in a few weeks. The first was the open model behind the Perplexity Decisions API, which makes yes/no and pick-one calls. This one is for search.

What an embedding model does

An embedding model turns text or images into lists of numbers so software can measure how close two things are in meaning. It's the part of a "chat with your documents" tool you never see. You ask a question, the system pulls the chunks of your files that look most relevant, and a chatbot answers from them. If retrieval grabs the wrong page, you get a confident answer from the wrong page.

Most embedding models compress a whole passage into a single vector. Late interaction, the approach Perplexity uses here, keeps a small vector for every token instead. When you search, each word in your question is matched against its closest partner in the document and the scores are added up. That's slower and heavier, but it holds on to detail a single vector smears out, like a part number buried in a long spec sheet.

What's new in this release

From the model card and the announcement:

• You can feed it rendered PDF pages and slides as images, so scanned forms, tables and charts can skip OCR. You can't mix text and images in one batch, though, so each type gets encoded separately.
• The two sizes share one embedding space. Build your index once with the 9B model, then run everyday queries with the 0.6B model on cheaper hardware against that same index.
• Both are built on Qwen3.5 and were distilled from a larger internal Perplexity model.

On the card's benchmark, ViDoRe v3, the 9B model scores 65.2% on page images and the 0.6B scores 62.3%. So the small one gives up about three points. Launch coverage also quotes a 92.4% score on the MADQA document-question benchmark for the 9B. That figure comes from Perplexity's own testing and isn't on the model card, so I'd file it as a vendor claim for now.

Is pplx-embed-v2-late free?

Yes. MIT is a permissive license: you can use the models, modify them and sell products built on them. Some recent "open" releases cap commercial use by company revenue, and this one has no such cap.

You pay in other ways. The Hugging Face page says no inference provider serves the model yet, and Perplexity's paid embeddings API still lists only its older single-vector models. That means a GPU server and a developer who's comfortable with Python and a recent version of the sentence-transformers library.

Storage is the other bill. One vector per token adds up quickly. An analysis of comparable late-interaction models found their indexes ran roughly 20 to 64 times the size of a standard index before compression. Those numbers come from similar models, not this one, but expect the same direction. Your vector database also has to support multi-vector search, which Qdrant, Weaviate, Vespa, LanceDB and Milvus do.

Our read

For anyone outside a research lab, the useful part is skipping OCR. Most small businesses keep what they know in messy PDFs: scanned contracts, equipment manuals, insurance forms, old invoices whose tables come out garbled when you extract the text. A retriever that looks at the page the way a person does should do better on exactly that kind of pile.

Whether to self-host it now depends on volume. With a few thousand pages, a hosted single-vector service is cheaper, easier and probably good enough. A firm with years of scanned paperwork that staff dig through by hand every week now has a free, commercially usable option worth testing.

Before choosing a model, work out which documents your team searches most and how long it takes them. If you're not sure where those hours go, our free process audit can help you find out.