Perplexity AI has released two new embedding models, pplx-embed-v2-late, featuring a 0.6B edge-optimized version and a 9B parameter model. Both models are available on Hugging Face under the MIT license, with a hosted API endpoint currently in development.
The 9B model achieved a 92.4% score on the MADQA benchmark. In the BrowseComp+ benchmark, it outperformed the leading dense model by 8.7 percentage points. Meanwhile, the 0.6B model recorded a 61.2% score on the ViDoRe v3 Markdown benchmark, securing the second-best result.
The models utilize a MaxSim scoring approach, maintaining a 128-dimension vector for every token rather than compressing documents into a single vector. This architecture allows pages to be encoded as images, eliminating the need for OCR. Both models were distilled from an 18B teacher model using LEAF-style training.
Technical requirements for deployment include sentence-transformers version 6.0.0 or higher and transformers version 5.4.0 or higher. The models are designed for CUDA GPU environments. While published checkpoints are in F32 format, the company estimates memory usage at 2 bytes per parameter for weights.
