Perplexity releases pplx-embed-v2-late, open embedding models for text and image search

Perplexity introduced pplx-embed-v2-late, a new family of embedding models, on October 7, 2026. The family comes in two sizes, 0.6B and 9B. According to Perplexity's announcement, the models can search both text and images, so PDF pages, slide decks and scanned documents can be searched directly without first being converted to text. The open weights are available on Hugging Face under the MIT license. A week earlier, the company released pplx-embed-v2-context-9b-preview, a preview model that encodes document chunks together with their surrounding context.

An embedding model turns a piece of text or an image into a vector, a list of numbers. Content with similar meaning gets similar vectors, so a search system can find documents that cover the same topic as a query even when they use different words. These models sit underneath search engines and RAG (retrieval-augmented generation), where an AI assistant first pulls relevant passages from a document collection and then bases its answer on them.

A separate vector for every token

Most embedding models compress a whole document into a single vector. According to Perplexity, this can lose information in long documents that cover several topics. pplx-embed-v2-late uses a late-interaction approach known as ColBERT: the model outputs a separate 128-dimensional vector for each token in the input. At search time, each query token is matched with its most similar document token and those scores are summed. This lets different parts of a query match different parts of a document.

Text and images searched with one model

The company says real document collections are made of PDFs, slide decks, screenshots, scanned pages, charts and tables. The usual pipeline first runs OCR to turn them into text, which can introduce errors and drops layout, table structure and figures. The new models process the rendered page directly so that text queries can search it. Perplexity adds that the models also support semantic search over natural images such as landscape photos. According to the announcement, the training data consists of 186 million query-document pairs drawn from 46 languages, and the company says it excluded datasets tied to the benchmarks it reports.

Both sizes share one embedding space

Both models are built on Qwen3.5 and were distilled from an internal 18B-parameter teacher model. Because they share one embedding space, an index built with the larger model can be queried with the smaller one. In the setup Perplexity suggests, a document collection is indexed once with the 9B model, while live user queries are encoded with the faster small model. The small model is designed to run on edge devices as well; according to the company, about 240 million parameters are active when it encodes text and about 340 million when it encodes images. This allows private documents to be encoded locally on a device and used alongside a large index hosted in the cloud.

In the comparisons Perplexity published, the 9B model leads the evaluated vision-language models on ViDoRe (V3), a widely used visual document retrieval benchmark, and the company says the 0.6B model matches models with about five times as many active parameters. The announcement also says the 9B model beat all compared models on BrowseComp+, a test in which an AI agent uses the embedding model as its search tool. Rival models lead on some image search tests.

Usage and API plans

Both models can be downloaded from Hugging Face and work out of the box with sentence-transformers version 6 and later; the model cards include Python usage examples. The Perplexity API Platform currently offers the first-generation pplx-embed-v1 models. The company says it will progressively roll out support for late-interaction, dense and contextual embeddings on the API Platform. A technical report with further details and complete evaluation results is planned for later this year.

The pplx-embed-v2-context-9b-preview model released last week addresses a different problem in RAG systems: context is lost when documents are split into chunks. A document is passed to the model as a list of chunks, the chunks are encoded together, and each chunk's vector reflects its surroundings. According to Perplexity, the model is trained to retrieve the chunk that contains the answer along with the supporting chunks needed to understand and verify it. It is also MIT licensed, with model files on Hugging Face.

Because late-interaction models store many vectors per document, they need more storage and a different indexing setup than single-vector models; Perplexity notes that this cost grows with document length. According to the model card, text and images cannot be mixed in the same call. The training data spans 46 languages, but no separate results are given for individual languages, so it is worth testing the models on your own examples before relying on them for non-English documents. Since the contextual model is a preview, its model card advises against mixing its embeddings with those from future releases. For a fuller picture, follow evaluations from independent researchers as well.

Sources: Perplexity: Multimodal embeddings beyond a single vector, Perplexity: Contextual embedding beyond the gold passage, Hugging Face: pplx-embed-v2-late-9b, Hugging Face: pplx-embed-v2-late-0.6b, Hugging Face: pplx-embed-v2-context-9b-preview.