Perplexity has open-sourced pplx-embed-v2-late, a pair of embedding models for multimodal search released on 7 October under an MIT licence on Hugging Face. Both models, a 9B version and a 0.6B version, share one embedding space, which is the detail that makes them interesting together rather than separately.
The intended workflow splits the work by location. The larger 9B model runs in the cloud and builds a search index over large datasets. The smaller 0.6B model then runs on edge devices, querying that same index without needing a server round-trip for every search. Because the two models share an embedding space, an index built once with the 9B model stays usable by the lightweight one.
On MADQA, a benchmark built from questions over 800 PDF documents, Perplexity reports 92.4% answer accuracy for the 9B model and 90.1% for the 0.6B model. For comparison, Mixedbread's Agentic Search scores 93.4% on the same test; Perplexity says that gap sits within the benchmark's confidence interval, effectively calling it a tie at the top.
A hosted Perplexity API for these models is planned but not live yet, and the company has not published a full technical report, so some details, including exact parameter counts, are still only visible on the Hugging Face model cards. For teams building retrieval over PDFs, scanned documents or mixed text-and-image data, a shared embedding space that scales from server to device is the more practical story here than the headline benchmark number.
Source: MarkTechPost
