Wisevoter
U.S.

Perplexity AI Released New Embedding Models

The company launched two open-weight embedding models capable of searching complex files without needing OCR.

Updated on Oct. 7, 2026 in Artificial Intelligence

Bold vector editorial illustration of translucent prisms creating a geometric lattice, representing advanced data searching architecture.
Perplexity AI has launched two new open-weight embedding models, enabling efficient document searching across text and images without traditional optical character recognition. AI Illustration. Upload story photo >

Live Poll

Do you believe open-source AI tools make developing new software easier for you?

Perplexity AI has released its new pplx-embed-v2-late models, a pair of open-weight embedding models available on Hugging Face. These tools leverage a ColBERT-style late-interaction architecture to search text, images, and slides without relying on traditional text parsing or OCR.

Why it matters

By introducing cross-model querying capabilities, these models allow users to index large document collections with a powerful model while performing searches with a smaller, more cost-effective model. This architecture significantly reduces compute costs for enterprises managing massive datasets.

The new models include 0.6B and 9B parameter versions that utilize 128-dimensional vectors for each token. Training was conducted using 186 million query-document pairs across 46 languages, with the larger model showing notable performance gains on the ViDoRe v3 benchmark.

The players

Perplexity AI

Perplexity AI is a company that develops AI-powered search and research tools.

Hugging Face

Hugging Face is a collaborative platform for hosting and sharing machine learning models and datasets.

The details

The models were trained using distillation from an 18B parameter teacher model and use MaxSim scoring to handle search interactions. By maintaining 128-dimensional vectors, the system can interpret visual content and complex document formats natively.

Timeline

  1. Perplexity AI released the models on October 7, 2026.

  2. Perplexity previously introduced its dense v1 and contextual v2 models in 2026.

The Tech Race

The transition toward late-interaction architectures like the ColBERT model represents a move away from legacy embedding systems that struggled with complex, multi-modal documents. This positions Perplexity AI as a direct competitor to other providers seeking to reduce high operational costs in large-scale vector search.

Developers and software engineers can now integrate cheaper and faster search functionality into applications requiring the analysis of PDFs, slides, and images. Users will likely benefit from improved search accuracy and faster response times when accessing internal company document databases.

The takeaway

The use of cross-model querying suggests a future where high-performance retrieval does not require massive compute power for every single search interaction. Organizations should look to these open-weight models as a way to lower the barrier to entry for proprietary document search engines.

Further reading

For more background on the sector, visit the Artificial Intelligence section.

Live Poll

Do you believe open-source AI tools make developing new software easier for you?