Perplexity's text-only 9B embedding model beats leading retrievers on visual documents

AIPerplexity says its 9B embedding model, trained only on text, outperforms the leading dense retrievers on visual documents. Using two 1024-dimensional embeddings per page, it beats Nemotron-3.5-Embed-8B on ViDoRe-V3 with a 4x smaller index.

Perplexity DevelopersOfficial

@perplexitydevs

View on X

Part of a thread · earlier post

The 9B was trained only on text, yet it outperforms the leading dense retrievers on visual documents.

With two 1024d embeddings per page, it beats Nemotron-3.5-Embed-8B on ViDoRe-V3 with a 4x smaller index.

https://pplx.ai/hf-embed-v2-9b
https://pplx.ai/hf-embed-v2-0.6b

Open source

Source details

Perplexity Developers · x.comPublished Nov 2, 4:51 PM ET · added here 5:21 PM ET

Original title: The 9B was trained only on text, yet it outperforms the leading dense retrievers on visual documents.

Topics

Open-source ecosystem Model releases Evals & benchmarks