pplx-embed-v2-late
An open-source multi-vector embedding model released by Perplexity that supports a unified embedding space for text and images. It employs a late interaction approach and is available in 9B and 0.6B parameter sizes, supporting multimodal retrieval and OCR-free PDF page retrieval.
Timeline (last 90 days)
Perplexity AI released the pplx-embed-v2-late series of models, comprising two parameter versions, 0.6B and 9B.
Both models have been made public on the Perplexity AI organization page on Hugging Face.
The report does not give specific benchmark names or numerical comparisons in the abstract; the conclusion of 'SOTA within the same size category' is currently based only on the official statement and still lacks third-party reproduction and verification.
The official technical report states that both models achieve SOTA on text, image and visual document retrieval benchmarks within their respective size categories.
pplx-embed-v2-late adopts a late-interaction architecture, retaining token-level representations and performing interaction computation at the retrieval stage.
Both pplx-embed-v2-late models were trained from an 18B teacher model via representation distillation.
pplx-embed-v2-late supports unified retrieval across text, images, and pages, sharing a single embedding space to enable cross-modal queries.
Perplexity released two embedding models named pplx-embed-v2-late, adopting a late-interaction architecture.
Both pplx-embed-v2-late models have been publicly released on Hugging Face with open weights.
Perplexity officially claims that both models achieve frontier performance, but without accompanying specific benchmark data or competitor comparisons.
10 more timeline events
All Facts (20)
The official technical report states that both models achieve SOTA on text, image and visual document retrieval benchmarks within their respective size categories.
50%UnverifiedThe report does not give specific benchmark names or numerical comparisons in the abstract; the conclusion of 'SOTA within the same size category' is currently based only on the official statement and still lacks third-party reproduction and verification.
50%Unverifiedpplx-embed-v2-late adopts a late-interaction architecture, retaining token-level representations and performing interaction computation at the retrieval stage.
50%UnverifiedBoth pplx-embed-v2-late models were trained from an 18B teacher model via representation distillation.
50%UnverifiedPerplexity officially claims that both models achieve frontier performance, but without accompanying specific benchmark data or competitor comparisons.
50%Unverifiedpplx-embed-v2-late supports unified retrieval across text, images, and pages, sharing a single embedding space to enable cross-modal queries.
50%UnverifiedMultimodal models capable of directly embedding images and pages can skip OCR or parsing conversion steps, offering practical value for building multimodal RAG systems.
50%UnverifiedDevelopers can build multimodal RAG systems, document search engines, or knowledge base Q&A applications based on pplx-embed-v2-late.
50%Unverifiedpplx-embed-v2-late is available in two versions, 9B and 0.6B, with weights published on Hugging Face.
50%UnverifiedThe 9B large model is responsible for indexing and building databases of multimodal data, while the 0.6B small model is positioned for on-device querying and can run directly on user devices.
50%UnverifiedThe two models share the same embedding space, so indexes built with the 9B model can be directly retrieved by query vectors from the 0.6B model.
50%Unverifiedpplx-embed-v2-late can retrieve PDF pages without OCR by processing PDF pages directly as images and mapping them into the same embedding space as text.
50%Unverifiedpplx-embed-v2-late achieved 92.4% on MADQA and 64% on BrowseComp+.
50%UnverifiedThe aforementioned benchmark results come from vendor self-testing and lack third-party comparative evaluation.
50%Unverifiedpplx-embed-v2-late extends the multi-vector late interaction mechanism to image tokens, enabling local image regions to participate in fine-grained matching.
50%UnverifiedBoth models have been made public on the Perplexity AI organization page on Hugging Face.
50%UnverifiedPerplexity AI released the pplx-embed-v2-late series of models, comprising two parameter versions, 0.6B and 9B.
50%UnverifiedBoth pplx-embed-v2-late models have been publicly released on Hugging Face with open weights.
50%UnverifiedPerplexity released two embedding models named pplx-embed-v2-late, adopting a late-interaction architecture.
50%UnverifiedPerplexity open-sourced the multi-vector embedding model pplx-embed-v2-late, which unifies text and images into a single shared embedding space.
50%