Perplexity Releases Open pplx-embed-v2-late Models

The pplx-embed-v2-late family retrieves visual documents without OCR and aligns 0.6B and 9B variants into a single shared vector space.

Perplexity published its pplx-embed-v2-late embedding family on Hugging Face under the permissive MIT license. Built on the Qwen3.5 architecture, the two open late-interaction models scale at 0.6 billion and 9 billion parameters for document retrieval and retrieval-augmented generation (RAG) pipelines. The company announced the release on its engineering blog and official X account (@perplexity_ai).

Direct visual retrieval without OCR pipelines

Most search setups for documents still extract raw text with optical character recognition (OCR) before turning paragraphs into vectors. That pipeline stumbles on slide decks, earnings statements, and technical manuals packed with tables and diagrams. Perplexity’s new models process rendered page images directly alongside plain text, building representations that keep visual relationships and formatting intact without extra preprocessing steps.

Shared vector space for asymmetric search

The models employ late interaction, a technique popularized by Stanford’s ColBERT. Instead of forcing an entire page into a single dense numerical vector, the architecture produces 128-dimensional vectors for every token. Relevance scoring then uses a MaxSim operation, summing the highest similarity scores between each query token and the document tokens.

The standout technical decision here is cross-model compatibility. The 0.6B model (340 million active parameters) and the 9B model (7.4 billion active parameters) map into the exact same vector space. That allows an asymmetric setup: engineering teams can run heavy indexing jobs offline with the 9B model to preserve fidelity, then deploy the 0.6B model on the user-facing side to answer live queries with minimal latency and hardware footprint, all without rebuilding the index.

Distillation and framework compatibility

Both models were distilled from an internal 18-billion-parameter teacher model, which Perplexity kept proprietary. The 0.6B variant underwent full fine-tuning, while the 9B version combined LoRA adaptation with fine-tuning on its top eight layers. The weights integrate directly into standard Python tooling, supported through the MultiVectorEncoder class in sentence-transformers (version 6.0.0 and above) and transformers (version 5.4.0 and above).

Reported benchmarks and operational trade-offs

According to benchmarks published by Perplexity, the 9B model scored an nDCG@10 of 65.2% on images and 64.7% on markdown within the ViDoRe v3 visual retrieval benchmark. The 0.6B model scored 62.3% and 61.2% on the same tests. Across 72 domain-specific text retrieval tasks, the models recorded 81.3% and 78.0%, respectively.

Because these figures come directly from the creator’s internal evaluation, real-world behavior on messy enterprise data still requires independent testing. Late-interaction systems also carry higher storage requirements than single-vector models, since storing multiple vectors per page demands more memory planning. Perplexity noted plans to bring the models to its commercial API platform, but did not announce a rollout timeline or pricing structure.