The models, available in 9B and 0.6B parameter versions, scored 92.4% on MADQA and 64% on BrowseComp+. Across 72 text retrieval tasks, the 9B model achieved 81.3 nDCG@10, while the 0.6B version reached 78.0. Perplexity reports that pairing 0.6B queries with 9B indexing improves text retrieval to 79.6 nDCG@10, and the 0.6B model's performance on ViDoRe v3 page images is within 1.2 points of the Nemotron ColEmbed V2 8B.
Tuesday, October 20, 2026
Story
Perplexity Open Sources 9B and 0.6B Embedding Models for OCR-Free Search
AI13 days ago
Perplexity has launched a set of late-interaction embedding models that allow for the retrieval of text and images within a single shared vector space. The new pplx-embed-v2-late weights are hosted on Hugging Face and enable users to search across PDF pages, screenshots, and slide decks as images, eliminating the need for optical character recognition. The framework is natively supported in Sentence Transformers and facilitates cross-model querying, where data is indexed by a larger model and queried by a smaller one to reduce latency.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.