Perplexity AI has launched two open-weight embedding models, pplx-embed-v2-late, available on Hugging Face. The models are offered in 0.6B and 9B parameter versions and are designed for multimodal search across text, images, PDFs, and slide decks without the need for OCR.
The models utilize a ColBERT-style late-interaction architecture, maintaining a 128-dimensional vector for each token. This approach allows for direct processing of complex layouts, tables, and figures. The training dataset comprised 186 million query-document pairs spanning 46 languages.
A key feature is the ability to perform cross-model querying. Users can index data using the 9B model and execute live searches with the smaller 0.6B model, which helps optimize cloud computing costs. The 9B model achieved an nDCG@10 score of up to 81.3% in specific tasks.
Benchmark results on ViDoRe v3 showed an improvement from 62.3% to 63.5%. These models expand Perplexity’s existing lineup of open-weight embeddings, which previously included the dense pplx-embed-v1 and contextual v2 versions released earlier in 2026.
