Key Info

Perplexity released pplx-embed-v2-late on Hugging Face under the MIT license, in 0.6B and 9B sizes built on Qwen3.5.

Highlights

  • ColBERT-style late-interaction retrievers: each token gets its own 128-dimensional vector, scored via MaxSim matching.
  • Handles text, images, and rendered pages such as PDFs and slides, potentially skipping OCR preprocessing.
  • Both sizes share one embedding space, so the small model can query an index built with the 9B.
  • Benchmarked on the public ViDoRe v3 image track (details truncated in source).