AI news · October 10, 2026
Perplexity releases multimodal late-interaction embeddings in 0.6B and 9B sizes
Perplexity's PPLX-Embed V2 Late emits 128 dimensions per token for text, images, and visual documents; the 0.6B encoder can query a 9B index, so retrieval can trade speed for quality.
- small model
- 0.6B
- large model
- 9B
- dimensions per token
- 128
TLDR AI's Oct 8 issue highlighted Perplexity's PPLX-Embed V2 Late, and the Hugging Face model card now provides the weights and usage code. The family uses late interaction: it keeps 128 dimensions for each token and scores query-document matches with MaxSim across text, images, and visual documents. The 0.6B and 9B models share an embedding space, so a smaller query encoder can search an index built with the larger model.
The model card lists MIT licensing, 340M active parameters for the 0.6B model, and 7.4B active parameters for the 9B model. On ViDoRe v3, the published image scores are 62.3% and 65.2%; the Markdown scores are 61.2% and 64.7%. It runs through Sentence Transformers, but no inference provider is listed, and mixed text-plus-image batches need separate encoding calls.
What you can do with it
Download the 0.6B or 9B MIT model from Hugging Face and run it with sentence-transformers. Use 0.6B for query encoding against a 9B index, then compare retrieval quality, storage, and latency on your own PDF and image set before replacing a single-vector search.
Our take
This matters for search over PDFs and images because the representation preserves token-level detail instead of one vector per document. The catch is operational: multi-vector retrieval costs more storage and scoring work, and the published benchmark is not evidence that it wins on your domain.
Links Download the modelTLDR coverage
Source: Hugging Face / Perplexity — Made With Models writes the brief; the reporting is theirs.