Perplexity released pplx-embed-v2-late on October 7, a pair of ColBERT-style embedding models for retrieving text, images and rendered document pages such as PDFs and slides from a single shared space. Both sizes are on Hugging Face under the MIT licence; a hosted API is planned but not yet live.

The line-up splits the work: a 0.6B model is built for fast, cheap live queries, including on edge hardware, while a 9B model is built for index quality. Because they share an embedding space, a corpus indexed with the 9B model can be searched with the 0.6B model, avoiding 9B inference cost on every query.

Instead of one vector per document, the models keep a compact 128-dimension vector per token and score matches with MaxSim. Perplexity reports 92.4% on the MADQA document-question benchmark for the 9B model, and the approach avoids an OCR step by encoding pages as images.