Perplexitypplx-embed v2
pplx-embed-v2-late-9b
This multimodal late-interaction model retrieves text, images and visual documents.
What the release provides
It adapts Qwen3.5 with bidirectional attention and teacher distillation. Each token has 128 dimensions and uses MaxSim matching. The two model sizes share a vector space. MIT terms apply.pplx-embed-v2-late-9b official model card
How to use it well
Practical guidance. Compare a smaller query model against the same larger-model document index. Measure retrieval quality before assuming the shared vector space removes every quality difference.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | perplexity-ai/pplx-embed-v2-late-9b |