Perplexity AI Releases pplx-embed-v2-late: A 0.6B Edge Model and a 9B Model Scoring 92.4% on MADQA

Perplexity has released pplx-embed-v2-late, a pair of ColBERT-style multimodal embedding models in 0.6B and 9B sizes, according to MarkTechPost. Both retrieve text, images and rendered PDF pages, and they share one embedding space. That means documents can be indexed with the large model in a datacenter and searched with the small one on a laptop or edge device.

The weights are on Hugging Face under the MIT license, so commercial use is allowed. A hosted Perplexity API is planned but not live. Perplexity reports the 9B model scored 92.4 percent on MADQA, a benchmark for agentic question answering over PDFs, with the 0.6B model at 90.1 percent.

We like the design choice here more than the headline number. Most search systems force a single trade-off: a big, expensive model for quality or a small, cheap one for speed. A shared embedding space lets teams put the heavy lifting at index time and keep queries fast and local. MarkTechPost reports that querying a 9B index with the 0.6B model recovered about half of the quality gap on text compared with using the small model on both sides. For privacy-sensitive document search, keeping the live query on device is a genuine benefit.

Treating pages as images, so no OCR step is needed, is also practical for anyone wrangling scanned reports and slide decks. And an MIT license on both sizes is generous compared with some rivals' non-commercial terms.

Now the caveats, which MarkTechPost lists plainly. All scores are self-reported and the technical report is not out yet. Storing a vector per token means index size grows with document length, which gets expensive at scale. It is not the leader on the ViDoRe v3 image retrieval benchmark, and a single input cannot mix text and images. None of that is disqualifying. It just means the numbers are a starting point, not a verdict.

Our read: a thoughtful open release that could make strong document retrieval cheaper to run privately. I'm watching for the technical report and independent benchmark runs before anyone builds procurement decisions on the 92.4 figure.

Dense embedding models compress a document into a single vector. Late-interaction models like this one keep a small vector for each token and match query tokens to their best document tokens, which tends to improve precision at the cost of storage. Perplexity distilled both sizes from a larger teacher model, per the report.