pplx-embed-v2-late pplx-embed-v2-late is a pair of late-interaction (ColBERT-style) embedding models from Perplexity, sized 0.6B and 9B, that retrieve across text, images, and rendered document pages in a single shared embedding space. The models search PDFs, slides, and scans without OCR, and the smaller model can query an index built by the larger one.
Visit productFeatured media
Loading media…
View on XCosigns Received
“Introducing the first multimodal embedding models from @perplexity_ai. Strong visual document retrieval, text-text retrieval, cross-model shared embedding space and many more cool capabilitie! 🥳”
“New company, new team, same frontier results & same open license Happy to announce the release of our new embedding model family More than just an upgrade in performance, the models are now multimodal, multi-vectors, and use a shared embedding space to allow cross-model querying”
“another release from us, we probably built the strongest late interaction retriever you can find ever, anywhere on earth. we trained a massive 18B colbert model for text retrieval, image retrieval and document screenshot retrieval, then distilled it to a 9B colbert, and a 0.6B colbert. Since they from the same teacher, so naturally have shared token embedding space. So you can use 9B offline indexing, 0.6B online search, or use 9B on strong chips and 0.6B locally. Remind 0.6b includes a vision tower as well! pplx-embed-v2-context, late both out, Dense will be soon. These will marks our new gen or retriever suite.”
“Please welcome pplx-embed-v2-late! Extra strong late interaction models (9B and 0.6B). As usual, available on @huggingface :) Kudos to @bo_wangbo , @antoine_chaffin and team! https://www.perplexity.ai/hub/blog/multimodal-embeddings-beyond-a-single-vector”
“We’re open-sourcing pplx-embed-v2-late, multi-vector embeddings for text and images, 9B and 0.6B, in one shared embedding space. You can use these to index multimodal data with 9B, and query on device with 0.6B. This also enables you to search over PDF pages with no OCR. And scores 92.4% on MADQA, 64% on BrowseComp+. Weights available on @huggingface now.”
“Perplexity's new retrieval models can search visual pages, including slides and scans. For company knowledge tools, that's useful: the answer may live in an old diagram or table. Making it findable for someone who doesn't know the filename could unlock a lot of buried work.”
“@perplexity_ai just put out new embedding models called pplx embed v2 late and made them free on Hugging Face. My take is this matters more than another chatbot update. Search and retrieval is the real backend every AI product needs. The clever part is you index with the big 9B model and search with the tiny 0.6B model and still get strong results. That cuts compute cost a lot for anyone running search at scale. If you build or sell AI agents and AI search for clients, swapping in a stronger embedding model like this is a fast upgrade you can bill for right now. Teams that only pitch GPT or Claude wrappers are skipping the part that actually makes search good, which is the embedding layer underneath.”