Perplexity AI has introduced pplx-embed-v2-late, a pair of embedding models offered in two configurations. The smaller 0.6B-parameter model is designed to run on edge devices, while the larger 9B-parameter model targets high-quality index construction.
According to the announcement, the models' strongest result is 92.4% on the MADQA benchmark, with the weakest reported score being 61.2% on ViDoRe v3 Markdown. The source does not provide additional benchmark context or comparisons to other embedding models.
Both variants are released under the MIT license and are ready for self-hosting. Since this is a single source, there are no differing viewpoints to note; the report is a straightforward product release summary.