Embed v1 4b

Embed v1 4b is the quality-focused tier of Perplexity's pplx-embed family, designed specifically for real-world web-scale text retrieval. Within that family it sits at the larger scale, built to maximize retrieval accuracy rather than lightweight embedding generation, and it is positioned alongside a companion pplx-embed-context-v1 line that emphasizes document-level surrounding context. The model is part of a generation that, according to Perplexity's research release, leads a range of public benchmarks including MTEB Multilingual, BERGEN, ToolRet, and ConTEB, and also delivers best-in-class results on Perplexity's internal web-scale evaluations PPLXQuery2Query and PPLXQuery2Doc.

In practical use, the model produces 2560-dimensional vectors that are quantized natively to INT8, with the option of binary quantization as well, allowing large reductions in storage overhead compared to FP32. It uses mean pooling, accepts inputs up to the cataloged API limit tokens without requiring an instruction prefix, and supports Matryoshka representation learning so shorter vectors can be requested when search speed and storage matter more than peak recall. This combination of strong benchmark performance, long context handling, no-prefix convenience, and compact INT8 output makes the 4B tier well suited to production retrieval pipelines, semantic search backends, and web-scale indexing workloads where embedding quality is the priority.