Perplexity AI has released pplx-embed-v2-late, a new set of embedding models comprising two distinct sizes. According to MarkTechPost, the release includes a 0.6 billion parameter model designed for edge devices and a larger 9 billion parameter model intended for constructing high-quality search indexes. Both versions are available under an MIT license, allowing developers to self-host the technology.
Performance and Capabilities
Embedding models translate text into numerical representations used for search, retrieval, and information matching. In benchmark evaluations, the models reached a high score of 92.4 percent on the MADQA dataset. Their lowest performance was recorded at 61.2 percent on the ViDoRe v3 Markdown benchmark. These figures highlight a stark performance variance depending on the specific evaluation task, showing strong capability in certain specialized question-answering contexts while facing more friction in structured document tasks like Markdown retrieval.
Unknown Factors and Future Metrics
Despite the public release, several details regarding the rollout remain absent. The announcement did not specify training dataset compositions, hardware requirements beyond edge deployment suitability, or exact latency metrics for the 0.6B edge variant. Developers looking to adopt the MIT-licensed weights will need to run their own local benchmarks to determine how the 9B and 0.6B versions handle specific domain-specific corpora outside of the standard academic test suites.
Perplexity AI’s latest release broadens the available options for teams seeking self-hosted retrieval infrastructure.
Lead image: Vladimir Srajber / Pexels
