We're releasing pplx-embed-v2-late, two late-interaction embedding models that retrieve text, images, and pages with a shared embedding space for cross-model querying.
Both models achieve frontier performance and are publicly available on Hugging Face.
perplexity.ai/hub/blog/multi…
Both models achieve frontier performance and are publicly available on Hugging Face.
perplexity.ai/hub/blog/multi…
Oct 7, 2026 · 4:09 PM UTC
72
120
682
123,800
Dense embedding models compress each document into one vector, which loses detail as pages get longer or more visual.
pplx-embed-v2-late keeps a 128-dimensional vector per token and scores with MaxSim, so each query token is matched to its closest token in the document.
pplx-embed-v2-late keeps a 128-dimensional vector per token and scores with MaxSim, so each query token is matched to its closest token in the document.
3
5
43
9,782
The models also embed images and rendered pages directly, so PDFs, slides, and scans are searched without OCR, keeping tables, figures, and layout that text extraction drops.
2
1
26
3,123
Both sizes are distilled token by token from one 18B teacher, so they share one embedding space.
A corpus indexed with the 9B model can be searched with 0.6B queries. That lifts ViDoRe v3 from 62.3% to 63.5% with no added query cost.
A corpus indexed with the 9B model can be searched with 0.6B queries. That lifts ViDoRe v3 from 62.3% to 63.5% with no added query cost.
2
1
16
2,424
On Q2D-Web, about 70,000 production-derived search queries over 190M web documents, the 9B and 0.6B models reach 74.8% and 73.6% Recall@1000, against 69.3% for nemotron-embed-8b.
4
0
16
2,439
In agentic search, a GPT-OSS-120B agent using the 9B model answers 64.0% of BrowseComp+ questions correctly, 4.9 points ahead of the next ColBERT model, with fewer searches than any baseline.
1
1
13
2,400
On MADQA, 500 questions over 800 PDFs, it reaches 92.4%, the top result among retrievers.
1
0
15
5,975
Both models are on Hugging Face: huggingface.co/collections/pe…
0
0
20
5,843
Perplexity just released pplx-embed-v2-late, a multimodal embedding family in 0.6B and 9B sizes that share 1 embedding space. A tiny 0.6B query encoder can search an index built by the 9B model.
The core idea is late interaction: keep a 128-dim vector for every token instead t, then score with MaxSim. It searches rendered PDF pages directly, with no OCR step.
The 9B tops BrowseComp+ at 64.0%, 4.9pp above the next ColBERT model, and hits 92.4% on MADQA. The 0.6B activates only 340M params for images and lands within 1.2pp of NVIDIA's 8B Nemotron ColEmbed on ViDoRe v3. The trade-off: index size grows with document length, and Tencent's EVIE still leads visual retrieval.
Full analysis: marktechpost.com/2026/10/07/per…
Model: huggingface.co/collections/pe…
Technical details: perplexity.ai/hub/blog/multi…
@perplexity_ai @AravSrinivas
0
0
1
231
@perplexity_ai retrieval models are just search engines with a better résumé
0
0
3
272
@perplexity_ai will you have this on your API soon? would strongly consider using this serverless.
0
0
2
267
@perplexity_ai Is this easily finetunable? Like if I have normal question answer pairs?
0
0
0
397
@perplexity_ai A shared embedding space for text, images, and webpages could make multimodal search much more seamless and useful in real world applications
0
0
0
18
@perplexity_ai Search that handles text, images and pages together? Nice, and it's public.
0
0
0
47
@perplexity_ai Late interaction usually gets traded for index size. Curious what the storage and latency look like at 128 dims per token once a real corpus is indexed, since that is where dense models usually win the cost back.
0
0
0
18
@perplexity_ai i just watched a 0.6b model get promoted over my entire office
0
0
0
131
@perplexity_ai @bot look into this thread, do search on all the post covered in here and generate a new research note in our research database
1
0
0
30
@perplexity_aity 刚刚发布 pplx-embed-v2-late：一对 late-interaction 嵌入模型，在统一向量空间里检索文本、图片和页面，自称前沿性能，权重已在 Hugging Face 公开。
【Hroid观点】嵌入是 RAG 的地基。Perplexity 把 late-interaction（ColBERT 路线）做成产品并开源，直接和 Voyage、Jina、MixedBread 抢开发者心智。对开发者：图文混排文档的检索精度有望明显提升，开源权重意味着可本地部署、成本自己说了算。@perplexity_ai
#Perplexity #Embedding #RAG
【Hroid观点】嵌入是 RAG 的地基。Perplexity 把 late-interaction（ColBERT 路线）做成产品并开源，直接和 Voyage、Jina、MixedBread 抢开发者心智。对开发者：图文混排文档的检索精度有望明显提升，开源权重意味着可本地部署、成本自己说了算。@perplexity_ai
#Perplexity #Embedding #RAG
0
0
1
121
@perplexity_ai @bot how is this relevant to my work? send to my grok bot agent
2
0
1
56
@perplexity_ai The shared embedding space for cross-model querying is actually a very smart design choice. x.com/AGTPinsights/s…
Perplexity just open-sourced pplx-embed-v2-late, two multimodal search models. Here's what you need to know.
The release is two late-interaction embedding models for text and images, at 9B and 0.6B parameters. Both share one embedding space and are on Hugging Face under an MIT s both reach frontier performance.
Dense models squeeze each document into one vector, which loses detail on long or visual pages. These keep a 128-dimensional vector per token and score with MaxSim, matching each query token to its closest document token.
They also embed images and rendered pages directly. PDFs, slides and scans are searched with no OCR, so tables, figures and layout are kept.
Both sizes are distilled token by token from one 18B teacher, which is why they share a space. You can index with the 9B and query on device with the 0.6B. Perplexity's CEO Aravind Srinivas pitched exactly that setup. Researchers like Omar Khattab and Tony Wu called the models very strong and said cross-model retrieval cuts latency and helps edge devices.
Key numbers:
- ViDoRe v3 page images: 0.6B scores 62.3 nDCG@10, 9B scores 65.2
- 0.6B is within 1.2 points of Nemotron ColEmbed V2 8B
- 0.6B queries on a 9B index lift ViDoRe v3 from 62.3 to 63.5, at no added query cost
- Q2D-Web, about 70,000 production-derived queries over 190M web documents: 9B hits 74.8% Recall@1000, 0.6B hits 73.6%
- nemotron-embed-8b scores 69.3% on the same test
Install with sentence-transformers 6.0.0 or newer and transformers 5.4.0 or newer. Perplexity published the details in an Oct 7, 2026 blog post.
0
0
0
158
@perplexity_ainegocio chico que guarda catálogos y fichas técnicas puede aprovechar esto: los modelos pplx-embed-v2-late de Perplexity recuperan texto, imágenes y páginas en un espacio compartido de embeddings, para búsqueda entre modelos, y el anuncio dice que ya están disponibles en Hugging Face 🔍
0
0
0
21
@perplexity_ai Late-interaction retrieval keeps more detail, at a real cost in index size and latency. For a RAG system the decision is memory budget versus recall, and the eval set should reflect your own documents, not a benchmark.
0
0
0
5
@perplexity_ai un vector por token suena bárbaro para PDFs largos con tablas. ¿cuánto más pesa el índice comparado con uno de un vector por documento?
0
0
0
54
@perplexity_ai Searching PDF pages with no OCR is the part I care about most. Scanned invoices and charts lose so much in OCR. Indexing with the 9B and querying on-device with the 0.6B is a nice split for private files too.
0
0
0
2
@perplexity_ai ファーストビューで印象を掴む
でも、その前に「情報の階層設計」が重要
視線誘導の設計がうまくいくと
ユーザーの脳は自然と重要情報に引き寄せられる
LPの構成って、味付けじゃなくて盛り付けなんだと思う
でも、その前に「情報の階層設計」が重要
視線誘導の設計がうまくいくと
ユーザーの脳は自然と重要情報に引き寄せられる
LPの構成って、味付けじゃなくて盛り付けなんだと思う
0
0
0
2
@perplexity_ai Late-interaction retrieval is such a fun rabbit hole, love the push on embeddings! Can't wait to see builders put these to work!
0
0
0
20
@perplexity_ai For the 9B index with 0.6B queries, do you have storage per rendered page after compression?
0
0
0
57
@perplexity_ai A lot of RAG misses come from tables and charts stuck inside PDFs. Curious how this handles a scanned spec sheet.
0
0
0
54
@perplexity_ai The shared space is the clever part: build the index with the 9B once, query it with the 0.6B (340M active). The cost moves to storage: one 128-dim vector per token, so a 1,000-token page is ~256 KB at fp16 before compression. One 1024-dim vector is 2 KB.
0
0
0
15
@perplexity_ai The 0.6B model matching 5x larger competitors on NDCG@10 charts is a significant efficiency win for practical deployment.
0
0
0
18
@perplexity_ai If you don't code, this won't be directly usable. The real question is which search or chat product ends up shipping it, and when.
0
0
0
182
@perplexity_ai Nice, 62.3 from 340M active params, matching models five times bigger.🎉
0
0
0
111
@perplexity_ai Late-interaction keeps a vector per token, not one per doc, so the index gets many times bigger. Check storage on your real corpus before swapping it in.
0
0
0
95