Updated

pplx-embed-v2-late (0.6B and 9B)

Perplexity released two late-interaction embedding models, pplx-embed-v2-late-0.6b and pplx-embed-v2-late-9b, for text and image retrieval. They retain token-level vectors and share an embedding space, and both are publicly available on Hugging Face.

By No Code MBA · Facts checked

Provider
Perplexity
Announcement date
Status at last check
Generally available
Model family
pplx-embed v2
Preceding model
pplx-embed-v1
Capabilities
Embeddings, Text, Image

What changed

These are ColBERT-style late-interaction models that keep a 128-dimensional vector per token and score with MaxSim, covering text and images such as rendered PDF pages in a shared embedding space. Both are distilled from an unreleased 18B teacher and built on Qwen3.5; the 0.6B model has about 240M active parameters for text. Perplexity reports ViDoRe v3 visual nDCG@10 of 65.2% (9B) and 62.3% (0.6B), and says an index built with 9B can be queried with 0.6B.

Access and availability

Open weights in the perplexity-ai/pplx-embed-v2 collection on Hugging Face under the MIT license, usable with sentence-transformers 6.0.0+ and transformers 5.4.0+. Perplexity says late-interaction embeddings will roll out progressively on its API Platform.

Official access documentation

What you can build

Compare the 0.6B and 9B variants on a representative set of text queries and scanned or rendered documents; measure retrieval relevance alongside index size and query latency.

Limitations

Late interaction stores a vector for every document token, so storage and scoring cost grow with document length. Mixed text and image inputs in one batch are not supported, and the full technical report and API availability were still pending.

Sources and release history