Perplexity open-sources a new retrieval model: its 0.6B small model can directly query indexes built by the 9B model.

Beating AI News: Perplexity has open-sourced two multimodal embedding models, pplx-embed-v2-late, with sizes of 0.6B and 9B parameters respectively. The two models are vector-compatible, allowing developers to use the larger 9B model for building knowledge bases while leveraging the smaller 0.6B model for daily searches—eliminating the need to run the large model for every query. The new models support text, image, and PDF page retrieval. Traditional embedding models typically compress content into a single vector, which often leads to detail loss. In contrast, the new models retain a 128-dimensional vector for each token, enabling search queries to match relevant segments within documents directly. When processing PDFs, PPTs, and scanned materials, the models can convert page images straight to vectors, bypassing OCR text extraction while preserving charts, tables, and layout information. Both models are trained from the same 18B teacher model and share a common vector space. In official ViDoRe v3 image retrieval tests, the 0.6B model scored 62.3% when used for both database building and queries; switching to the 9B model for database building and the 0.6B model for queries boosted the score to 63.5%; the full 9B model achieved 65.2%. The model weights are now available on Hugging Face under the MIT license.