pplx-embed-v2-late Technical Report

Markus Krimmel; Louis Milliken; Ekaterina Glazkova; Maximilian Schall; Antoine Chaffin; Sedigheh Eslami; Jonas Golde; Denis Bykov

cs.IROct 21, 2026

pplx-embed-v2-late Technical Report

Authors: Markus Krimmel, Louis Milliken, Ekaterina Glazkova, Maximilian Schall, Antoine Chaffin, Sedigheh Eslami, Jonas Golde, Denis Bykov,

Organizations: Perplexity AI

Abstract

In this report, we introduce pplx-embed-v2-late, a pair of 0.6B and 9B multilingual and multimodal late-interaction models. Both models achieve state-of-the-art results within their respective size classes across a range of text, image, and visual document retrieval benchmarks. On visual document retrieval tasks, pplx-embed-v2-late-0.6b is competitive with models several times larger than itself. The models are trained via representation distillation from an 18B teacher. Consequently, they share an embedding space that supports asymmetric deployment: documents can be indexed with the 9B model and queried with the low-latency 0.6B model. Both models are available at https://huggingface.co/collections/perplexity-ai/pplx-embed-v2.

Figures & tables

Figure 1: Comparison of our sampling method with independent random sampling according to the dataset weights.

Model	Dim.	Vectors	Avg.	Finance (18)	Legal (20)	Health (8)	Conv. (5)	Tech (4)	Multiling. (17)
>1B / size unreported
pplx-embed-v2-late-9b	128	multi	81.3	82.5	81.6	82.5	81.5	70.4	89.5
nemotron-3-embed-8b	4096	single	79.7	80.8	81.1	83.9	77.4	67.5	87.9
gemini-embedding-2	3072	single	78.3	77.0	78.5	82.4	76.2	67.7	88.1
voyage-4-large	1024	single	79.0	81.5	80.0	83.0	76.7	65.1	87.9
qwen3-embedding-8b	4096	single	74.8	72.3	76.8	80.2	72.0	62.7	84.7

Table 1: Domain-specific text retrieval performance (nDCG@10, %). The average weights the six domains equally; parentheses show the number of tasks in each domain. Dim. is the embedding dimension (per token for multi-vector models). Best score per column in bold.

>1B / size unreported
			ViDoRe-V3 (text)	Q2D-Web
Model	Dim.	Vectors	Avg.	Citation	Web	Combined
pplx-embed-v2-late-9b	128	multi	64.7	88.2	68.8	74.8
nemotron-3-embed-8b	4096	single	60.6	83.1	61.7	69.3
topk-embed-v1-2b	128	multi	59.2	79.2	57.7	65.3
gemini-embedding-2	3072	single	58.9	–	–	–

Table 2: Text retrieval on the public subset of ViDoRe-V3 after converting page images to Markdown with OCR (average nDCG@10, %) and on Q2D-Web (Recall@1000, %) under three relevance judgment sets. Dim. is the embedding dimension (per token for multi-vector models). topk-embed-v1 ViDoRe-V3 scores are taken from the model cards ( TopK, 2026a ; TopK, 2026b ) at 128 dimensions; on Q2D-Web, these models were evaluated with an output dimension of 128. – denotes not evaluated. Best score per column in bold.

Vision + language, >1B / size unreported
Model	Dim.	Vectors	ViDoRe-V3 (image)	MIRACL-Vision	PPLX-Q2I
pplx-embed-v2-late-9b	128	multi	65.2	75.7	65.0
topk-embed-v1-2b	128	multi	61.9	60.2	51.5
qwen3-vl-embedding-8b	4096	single	58.3	68.5	50.6
gemini-embedding-2	3072	single	46.0	77.5	67.0
Vision + language, <1B

Table 3: Text-to-image retrieval (nDCG@10, %) on the public subset of ViDoRe-V3 (average), MIRACL-Vision and PPLX-Q2I. Vision-only models are trained specifically for image retrieval and are listed separately from vision-language models. Dim. is the embedding dimension (per token for multi-vector models). Nemotron ColEmbed and EVIE models were not evaluated on PPLX-Q2I because their high output dimensionality makes indexing impractical; – denotes not evaluated. Best score per column in bold.

Figure 2: Cross-model retrieval nDCG@10 for three query/document model pairings. Results across six text domains.

Figure 3: BrowseComp+ accuracy and recall versus the average number of searches (GPT-OSS-120B agent, high effort).

Multimodal Embedding

Multimodal embedding maps text, images, audio, video, and other data into representations that support comparison and retrieval across modalities.

See all 159 papers

Can Frontier LLMs Match Natively Multimodal Embeddings? A Comparison on Hard-Negative Text-to-Image Retrieval
Aug 11, 2026Archan Dutta, Vyanktesh KanungoVLM Evaluation LLM Reranking
Westcliff University
PLUME: Latent Reasoning Based Universal Multimodal Embedding
Apr 2, 2026Chenwei He, Xiangzhao Hao, Tianyu Yang +6Multimodal Embedding Multimodal IR
Southeast University · Institute of Automation, Chinese Academy of Sciences · University of Chinese Academy of Sciences
Giga-Embeddings: Mixture-of-Experts Encoders for High-Throughput Text Embeddings
Aug 24, 2026Egor Kolodin, Egor Krasnoperov, Evgeniy Kosarev +1Text Embedding Models Mixture-of-Experts Models
1MIPT · 2SaluteDevices · 3MSU

Aug 11, 2026J/K move · Enter open · S save

Can Frontier LLMs Match Natively Multimodal Embeddings? A Comparison on Hard-Negative Text-to-Image Retrieval

Archan Dutta, Vyanktesh Kanungo

Westcliff University

VLM Evaluation LLM Reranking Multimodal Embedding Image Retrieval

Open paper AskPDF arXiv

Multimodal retrieval and classification across different types of media, spanning text, images,video and audio, has traditionally relied on dual-encoder models that align visual and textual representations through contrastive learning. The March 2026 release of Gemini Embedding 2, Google's first natively multimodal embedding model to map text, images, video, audio, and documents into a single shared space, raises competition among multimodal retrieval systems. Simultaneously, frontier Large language models (LLMs) have also demonstrated strong visual understanding, raising the question of whether they can serve as effective zero-shot rankers. Our study provides the first direct comparison of native multimodal embeddings against LLM-based visual ranking on Flickr30k. We observe that GPT-4.1 and Claude Sonnet 4.6 perform on par with Gemini Embedding 2. Additionally, once embeddings are precomputed, multimodal embeddings are better suited for low-latency applications.

Abstract

Multimodal Embedding

Explore similar work

Can Frontier LLMs Match Natively Multimodal Embeddings? A Comparison on Hard-Negative Text-to-Image Retrieval

PLUME: Latent Reasoning Based Universal Multimodal Embedding

Giga-Embeddings: Mixture-of-Experts Encoders for High-Throughput Text Embeddings