pplx-embed
Perplexity's MIT-licensed embedding family: dense, contextual and late-interaction models, from 0.6B to 9B, with images in v2.
In short. pplx-embed turns text, and in the newer versions images of documents, into numbers so that a search system can find the most relevant passages. Perplexity ships several kinds. Dense models give one vector per text, contextual models embed a chunk while seeing the rest of its document, and late-interaction models keep one small vector per token for finer matching.
Summary
Perplexity published pplx-embed-v1 on Hugging Face in January 2026 as text embedding models of 0.6B and 4B parameters, with a contextual variant, pplx-embed-context-v1, for document chunks in retrieval-augmented generation. They were built on Qwen3 models that Perplexity continued to pre-train with a diffusion objective. They give 1024 or 2560 dimensions, accept up to 32K tokens, support Matryoshka truncation and need no instruction prefix. They output unnormalized int8 vectors natively, with a binary option, so compare them with cosine similarity. The card calls them state-of-the-art for web-scale retrieval but gives no benchmark numbers. They are also available through Perplexity's embeddings API.
The v2 models, published from July to October 2026, add images. pplx-embed-v2-late is a family of multimodal late-interaction models in 0.6B and 9B sizes, built on Qwen3.5 with bidirectional attention. They give one 128-dimensional vector per token and score with MaxSim, so documents can be pages as images, and the small model can query an index built with the large one. Perplexity's own ViDoRe v3 image scores are 62.3% for the 0.6B and 65.2% for the 9B, as nDCG@10. pplx-embed-v2-context embeds chunks of text or images with their surrounding document in view, has 1024 and 2048 trained Matryoshka dimensions, and is described in arXiv 2610.38854, which says it was distilled from a query-aware context compression model. The weights take about four bytes per parameter, so the repositories are 2.4 GB, 16.1 GB and 33.6 GB.
The cards use the MIT license for the repositories we checked, but the pplx-embed-v2-context-0.6b card states no license. Queries and documents must be encoded with different methods in the v2 context models, chunks should not overlap, and they need custom code, Transformers 5.4 or later and, for speed, flash-linear-attention. In the 30 days to 11 November 2026, Hugging Face counted about 64,700 downloads of pplx-embed-v1-0.6B, 7,400 of the 4B and about 2,800 each of the contextual v1 0.6B and the v2 late 0.6B.
Good for
- Web-scale semantic search and retrieval-augmented generation with no instruction prefix to maintain
- Embedding document chunks with their surrounding context, in text or as page images
- Retrieving visual documents such as scanned pages and slides with late-interaction scoring
- Cheap storage and fast search with native int8 or binary vectors and Matryoshka truncation
Watch out for
- The v1 card claims state-of-the-art quality without numbers, and the v2 scores we quote are Perplexity's own
- The embeddings are unnormalized int8 vectors, so use cosine similarity, and the v2 context models need different calls for queries and documents
- Late-interaction models store one 128-dimensional vector per token, so the index is much larger than with a dense model
- The v2 models need custom code and a recent Transformers, and the pplx-embed-v2-context-0.6b card states no license
- The weights take about four bytes per parameter, so the downloads are large for the size
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of pplx-embed. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| pplx-embed-v2-late-0.6B (multimodal ColBERT) FP32 | 590M | 2.4 GB | 2.76 GB | Laptop or phone |
| pplx-embed-v1-0.6B (dense text, 1024 dimensions) FP32 | 600M | 2.4 GB | 2.76 GB | Laptop or phone |
| pplx-embed-context-v1-0.6B (contextual chunks) FP32 | 600M | 2.4 GB | 2.76 GB | Laptop or phone |
| pplx-embed-v1-4B (dense text, 2560 dimensions) FP32 | 4.02B | 16 GB | 19 GB | Gaming GPU |
| pplx-embed-v2-late-9B (multimodal ColBERT) FP32 | 8.39B | 34 GB | 39 GB | Workstation |
| pplx-embed-v2-context-9B (multimodal contextual) FP32 | 8.4B | 34 GB | 39 GB | Workstation |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
Frequently asked questions
- What is pplx-embed?
- pplx-embed turns text, and in the newer versions images of documents, into numbers so that a search system can find the most relevant passages. Perplexity ships several kinds. Dense models give one vector per text, contextual models embed a chunk while seeing the rest of its document, and late-interaction models keep one small vector per token for finer matching.
- Can I use pplx-embed commercially?
- Yes. pplx-embed is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is pplx-embed open source?
- pplx-embed is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run pplx-embed on my own computer?
- Yes. The smallest variant, pplx-embed-v2-late-0.6B (multimodal ColBERT), needs about 2.76 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was pplx-embed released?
- pplx-embed was released by Perplexity on January 14, 2026.
- How long a context window does pplx-embed have?
- pplx-embed supports a context window of up to 32K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- pplx-embed-v1-0.6b model card (Hugging Face) huggingface.co
- pplx-embed-v2-late-9b model card (Hugging Face) huggingface.co
- pplx-embed-v2-context-9b model card (Hugging Face) huggingface.co
- pplx-embed-v2-context-0.6b model card (Hugging Face) huggingface.co
- Contextual Embeddings Beyond the Gold Chunk (arXiv 2610.38854) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3-Embedding Alibaba Qwen | 600M – 7.57B | 32K | Apache 2.0 (Permissive) | 1.38 GB Laptop or phone | |
| EmbeddingGemma 2 Google DeepMind | 270M – 740M | 8K | Apache 2.0 (Permissive) | 0.621 GB Laptop or phone | |
| BGE-M3 BAAI | 568M | 8K | MIT (Permissive) | 2.61 GB Laptop or phone | |
| Jina Embeddings v5 Jina AI | 239M – 677M | 32K | CC BY-NC (Non-commercial) | 0.552 GB Laptop or phone |