最新模型
Perplexity Releases Multimodal Retrieval Models pplx-embed-v2-late, Enabling Small Models to Query Indexes Built by Larger Models
On October 7, Perplexity announced two open-weight late-interaction models for retrieving text, images, and visual documents. The 0.6B and 9B versions share an embedding space, allowing different-sized encoders to be used for document indexing and online queries, though their actual costs and performance on Chinese remain to be validated.
On October 7, Perplexity announced the release of pplx-embed-v2-late-0.6b and pplx-embed-v2-late-9b. The weights are available on Hugging Face under the MIT license. The release focuses on multimodal retrieval: text queries can be matched against text, images, and visual documents, and the two model sizes share an embedding space. Official announcement
The models are based on Qwen3.5 and use bidirectional attention. They produce a 128-dimensional vector for each token, then use MaxSim to calculate query-document similarity. This ColBERT-style late interaction retains token-level representations, so retrieval systems need to handle multi-vector data. The shared space allows documents to be pre-encoded with the 9B model and online queries processed with the 0.6B model. Its potential engineering value is that compute allocation for document encoding and query serving can be evaluated separately. 9B model card
According to the publisher's Public ViDoRe v3 evaluation, image retrieval nDCG@10 scores are 62.3% and 65.2%, while Markdown retrieval scores are 61.2% and 64.7%, respectively. These figures show how the two models compare on the specified benchmarks, but are not enough to infer performance on Chinese enterprise documents or to directly establish the quality and latency of using the smaller model to query an index built by the larger one. 0.6B model card
For deployment, the official example uses Sentence Transformers' MultiVectorEncoder and requires sentence-transformers >= 6.0.0 and transformers >= 5.4.0. Text and images must be encoded in separate batches; mixed inputs are not supported. When integrating with PyLate, teams should also verify the placement of the query and document markers. Chinese RAG teams should next evaluate cross-size query quality, index capacity, and end-to-end latency using data that includes tables, scanned pages, and mixed Chinese-English content, then determine whether the shared embedding space delivers a practical cost advantage. Deployment instructions