pplx-embed-v2-late-0.6b
Small late-interaction (ColBERT-style) embedder that searches text and PDF pages.
pplx-embed-v2-late-0.6b is a late-interaction embedding model from Perplexity: it keeps 128-dimension vectors per token and scores with MaxSim instead of one vector per input. It embeds text and images, so rendered PDF pages can be searched without OCR, and shares an embedding space with the 9B model so it can query a 9B-built index. Distilled from an 18B teacher (594M parameters).
- Provider
- Perplexity
- Type
- Embedding model
- Released
- Oct 7, 2026
- Input
- text, image
- Output
- embedding
- Open weights
- Yes
- License
- MIT
Best for
- Research
- Vision
- High volume / low cost
Strengths
- Searches PDF pages without OCR
- Shares an index with the 9B model
More from Perplexity
- Decider v1.1 27B — Perplexity. Perplexity’s open-weight multimodal decision model.
- pplx-embed-v1 4B — Perplexity. Perplexity’s web-scale retrieval embedding.
- pplx-embed-v2-late-9b — Perplexity. Perplexity’s large late-interaction embedder for text and visual document retrieval.