Perplexity AI Releases pplx-embed-v2-late: 0.6B and 9B Models
Perplexity launches pplx-embed-v2-late embedding models in 0.6B edge and 9B sizes, scoring 92.4% on MADQA under an MIT license for self-hosting.
Stock photo for illustration only, not from the actual event
- Perplexity releases pplx-embed-v2-late in 0.6B and 9B sizes
- The 9B model achieves a peak score of 92.4% on MADQA
- The 0.6B model is optimized to run on edge devices
- Both models are MIT-licensed and ready for self-hosting
Perplexity AI has announced the release of its latest embedding model lineup, named pplx-embed-v2-late. Developed to enhance data processing and high-performance indexing, the release introduces two distinct model sizes tailored for various deployment environments ranging from local hardware to powerful servers.
The lineup features a compact 0.6-billion parameter (0.6B) model specifically engineered to operate efficiently on edge devices, alongside a larger 9-billion parameter (9B) model optimized for constructing high-quality search and retrieval indexes.
In terms of benchmark performance, the models demonstrated strong capabilities, recording their highest score of 92.4% on the MADQA dataset, while their lowest score was 61.2% on the ViDoRe v3 Markdown benchmark.
The introduction of these embedding models highlights Perplexity's focus on making advanced AI infrastructure more versatile and accessible. Providing a lightweight 0.6B model for edge computing allows applications to process information locally, which helps reduce latency and enhance user data privacy.
Furthermore, both versions of the pplx-embed-v2-late models are released under an open MIT license and are fully prepared for self-hosting, allowing developers and organizations to integrate them directly into their own infrastructure.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment