pplx-embed-v2-late
Perplexity's late-interaction multimodal embedding models (0.6B and 9B) with 128-dim per-token vectors, open on Hugging Face.
Overview
Perplexity released pplx-embed-v2-late on October 7, 2026: ColBERT-style late-interaction embedding models in 9B and 0.6B sizes that output 128-dimensional vectors per token for text and images. Perplexity reports the 9B averages 81.3% nDCG@10 across 72 domain-specific text retrieval tasks and 64.7% on ViDoRe v3 Markdown, versus 60.6% for the next best external model. Both models are on Hugging Face and work with sentence-transformers 6.0.0+; API access for late-interaction, dense and contextual embeddings is planned later.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Late-interaction (multi-vector) retrieval with 128-dim per-token vectors
- 9B quality model and 0.6B edge-friendly model (about 240M active for text)
- Text and image retrieval in one model family
- A 9B index can be queried with the 0.6B model (63.5% on ViDoRe v3 image retrieval)
- Public weights on Hugging Face; Perplexity API rollout planned
Use cases
- Retrieval over visual documents such as PDFs and slides
- RAG pipelines needing higher-recall late-interaction search
- Cheap query-side encoding with the 0.6B model against a 9B-built index
Pricing
Free public weights on Hugging Face (license not stated in the announcement); Perplexity API pricing not yet published.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 7 Oct 2026
- Record updated
- 15 Oct 2026
- Record last checked
- 15 Oct 2026