pplx-embed

pplx-embed is a family of open-weight text embedding models developed by Perplexity for web-scale first-stage retrieval. The family includes two model types: pplx-embed-v1, designed for standard dense text retrieval, and pplx-embed-context-v1, which embeds passages with respect to surrounding document-level context (late chunking). Each model is available at 0.6B and 4B parameter scales.

Perplexity uses these models as the initial retrieval stage for its own search architecture, filtering documents from web-scale crawls before passing candidates to downstream rerankers and language models.

Model Specifications & Architecture

• Checkpoints: pplx-embed-v1-0.6B, pplx-embed-v1-4B, pplx-embed-context-v1-0.6B, and pplx-embed-context-v1-4B.
• Dimensions & Context: 1,024 dimensions for 0.6B models; 2,560 dimensions for 4B models. All variants support a 32,000-token context window, use mean pooling, and support Matryoshka Representation Learning (MRL).
• No Instruction Prefixes: Designed to embed inputs directly without requiring task-specific instruction prompts, eliminating indexing/query instruction mismatches during retrieval.
• Bidirectional Diffusion Backbones: Built from Qwen3 base models converted from causal decoders into bidirectional encoders via diffusion-based continued pretraining on 250 billion tokens across 30 languages.
• Native Quantization: Trained using quantization-aware training (tanh-based mean pooling and straight-through gradient estimation) to output native INT8 (4x storage reduction vs. FP32) and binary embeddings (32x storage reduction).

Vendor-Reported Benchmarks

In evaluations published by Perplexity:

• MTEB (Multilingual, v2 Retrieval): pplx-embed-v1-4B (INT8) achieves an average nDCG@10 of 69.66% (compared to 69.60% for Qwen3-Embedding-4B and 67.71% for gemini-embedding-001).
• ConTEB (Contextual Retrieval): pplx-embed-context-v1-4B (INT8) reaches an average nDCG@10 of 81.96% (compared to 79.45% for voyage-context-3 and 72.4% for Anthropic Contextual).
• BERGEN (End-to-End RAG): pplx-embed-v1-4B outperforms Qwen3-Embedding-4B on four of five evaluated question-answering tasks.
• Internal Web-Scale Evaluation: On PPLXQuery2Query (2.4M document corpus), pplx-embed-v1-4B achieves 73.5% Recall@10, and the 0.6B model reaches 71.1%. On PPLXQuery2Doc (30M document corpus), pplx-embed-v1-4B achieves 91.7% Recall@1000.

Deployment & Ecosystem

Model weights are published on Hugging Face under the MIT License and are served through the Perplexity API. Supported local and self-hosted inference runtimes include Hugging Face Transformers, SentenceTransformers, Text Embeddings Inference (TEI), and ONNX.