Perplexity has released two open-weight embedding models that can search across text and images, with the company saying they can also handle PDFs and slide decks without OCR preprocessing. Their shared embedding space means a collection indexed with the larger model can be searched using the smaller one, a useful bit of flexibility for teams weighing search quality against running costs.
Watch Desk analysis
What happened
The pplx-embed-v2-late family comes in 0.6-billion- and 9-billion-parameter versions. Perplexity says both support late-interaction, multi-vector representations and are available on Hugging Face, with compatibility for Transformers and sentence-transformers.
The company says the models can work across text, images, PDFs and slide decks without first converting image-based material through OCR. It also reports state-of-the-art results on several vision, text and agentic question-answering benchmarks; those are Perplexity’s performance claims, not independently established results here.
Why it matters
Embedding models help software find relevant material, rather than generate an answer directly. Supporting multiple formats could make one search system more useful across a mixed document collection, while the shared embedding space offers a way to use a smaller model for queries against data indexed with the larger one.
That combination gives developers something concrete to test: whether the convenience of searching across formats holds up on their own documents, and whether the smaller model is good enough for everyday queries. No need to pretend every PDF folder is about to become an oracle.
Our read
This is a worthwhile open-model release because the practical pitch is specific: cross-format search, two model sizes and a shared space between them. Developers building retrieval systems can try the models against their own collections and compare relevance and resource use before changing an existing setup. Benchmark superlatives can wait for outside testing.
What to watch
- How the models perform on independent evaluations and real document collections.
- Whether the smaller model can search larger-model indexes without a meaningful drop in relevance.
- What hardware and serving costs look like in practical deployments.
Discussion spark: Would you prioritise one shared embedding space across model sizes, or use separate models for indexing and search if that gave you better results?
Sources and evidence
- Perplexity: Perplexity releases pplx-embed-v2-late embedding models (7 October 2026, 16:09 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.