pplx-embed: State-of-the-art embedding models for web-scale retrieval

Perplexity has released pplx-embed-v1 and pplx-embed-context-v1, text embedding model families for standard and document-context-aware web-scale retrieval, each available in 0.6B- and 4B-parameter versions. The models use diffusion-based continued pretraining to convert Qwen3 causal language models into bidirectional encoders, followed by multilingual contrastive, contextual, and hard-negative training to improve query-document matching across 30 languages. They generate native INT8 and binary embeddings through quantization-aware training, cutting storage needs by fourfold and 32-fold respectively while retaining most retrieval quality, and they require no instruction prefixes. Perplexity reports that the models lead or match competing systems on benchmarks including MTEB, ConTEB, BERGEN, ToolRet, and internal web-scale retrieval tests, with the 4B contextual model achieving a reported state-of-the-art ConTEB score of 81.96% nDCG@10. The models are available under the MIT License on Hugging Face and through the Perplexity API, with support for several common inference frameworks.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.