Published: 8 October 2026
Perplexity released pplx-embed-v2-late on 7 October 2026: two open embedding models that search text, images and PDF pages in one index, without turning pages into text first. Both ship under the MIT licence, so you can run them commercially on your own servers. The small 0.6B model can query an index built by the 9B model, so the heavy work happens once and each search stays cheap.
Near the top on PDF questions, not first on page images
An embedding model turns documents into numbers so a search system can match them by meaning. These models keep one small vector per token instead of one per document, a design called late interaction. Perplexity’s results put the 9B model at or near the top on document and web search, with one clear exception.
| What is measured | pplx-embed-v2-late 9B | Best rival Perplexity names |
|---|---|---|
| MADQA (questions over 800 PDFs) | 92.4% | Mixedbread Agentic Search 93.4%, Mixedbread retriever 88.9% |
| Q2D-Web, Recall@1000 (finding web pages) | 74.8% | Previous best 69.3% |
| Domain text retrieval (72 tasks) | 81.3% | Next best 1.6 points lower |
| ViDoRe v3 visual (page images) | 65.2% | Only EVIE scores higher |
| BrowseComp+ (research agent) | 64.0% | Next best late-interaction model 4.9 points lower |
All results are Perplexity’s own, and the full technical report is promised for later this year. On MADQA, Perplexity says the gap with Mixedbread Agentic Search falls within its confidence interval.
There are limits. Storage grows with document length, because every token keeps a vector of 128 numbers. One input cannot mix text and images. The 9B model needs about 16 to 18 GB of GPU memory, by MarkTechPost’s estimate. A hosted Perplexity API is planned, with no date or price.
Test this if you search scanned contracts, reports or manuals where tables and layout matter and text extraction keeps failing. Self-hosting under MIT suits organisations that cannot send documents to an outside API. Perplexity trained on 46 languages but does not list Dutch, so try it on your own Dutch files before you rebuild an index. If plain-text search already works for you, there is little reason to switch.
For more information, visit the official announcement of pplx-embed-v2-late on the Perplexity blog.
