Perplexity Open-Sources pplx-embed-v2-late, 92.4% on MADQA
TL;DR
- The 0.6B model queries indexes built by the 9B model with no rebuild required, removing the need for separate vector stores per speed-quality tier.
- Both models embed PDFs, slides, and images natively without OCR, preserving table structure and layouts that text-extraction pipelines corrupt.
- The 9B scores 92.4% on MADQA but 65.2% on ViDoRe v3 image retrieval, marking a visible gap between text and visual document retrieval at current scale.
Perplexity has released two open-weight embedders, a 0.6B edge model and a 9B, that share one embedding space, so an index built with the big model can be served by the small one. Both landed October 7 on Hugging Face under the MIT license, announced on the company forum.
The pitch is retrieval over text, images, and rendered pages without a chunking or OCR pipeline in front. Perplexity's model card describes them as "late-interaction (ColBERT) retrievers for text, images, and visual documents, built on Qwen3.5 with bidirectional attention," each producing "one 128-dimensional vector per token" scored with MaxSim.
The 9B scored 92.4% on MADQA, a set of 500 questions over roughly 18,000 PDF pages, and 64.0% on BrowseComp+ paired with GPT-OSS-120B. AlphaSignal's writeup puts that 4.9 points ahead of competing ColBERT models. On Q2D-Web the company lists 74.8% Recall@1000 across about 190 million documents.
Both weights come from a bigger teacher. "Both models were distilled from an internal 18B ColBERT teacher trained on pair and triplet data. Distillation used a token-level LEAF-style objective," the model card says. The 0.6B was fully fine-tuned; for the 9B, the final eight transformer layers were fully fine-tuned while the remaining layers and the vision encoder were adapted with LoRA.
The 0.6B lands at 594M total parameters, with 240M for text and 340M for images. The 9B card lists 7.4B active parameters out of 8B. API access is announced as forthcoming; today the release is for self-hosters.
It arrives a day after Google's EmbeddingGemma 2 and alongside a run of retrieval and RAG work we have been tracking this week, including an Aalto result on inverting ColPali indexes.
What others are reporting
-
Hugging Face (Perplexity first-party model card) Read →
First-party model card with architecture details (ColBERT on Qwen3.5, bidirectional attention), LEAF distillation methodology, LoRA vs full fine-tune split, and working code examples.
The models produce one 128-dimensional vector per token and score query-document similarity using MaxSim. The 0.6B and 9B models share an embedding space.
-
AlphaSignal Read →
Puts a concrete scale figure on the release (190M docs), details the asymmetric query config for edge latency, and draws out implications for agentic retrieval pipelines.
Late-interaction models encode queries and documents independently, then compare their token-level representations during scoring.
-
MarkTechPost Read →
Surfaces the storage cost penalty of per-token 128-dim vectors, documents benchmark variance across document types, and frames edge device deployment as the primary use case.
users can query high-quality indexes built by the 9B model from resource-constrained edge devices using only the 0.6B model
-
Crypto Briefing Read →
Frames the cross-model query feature as an infrastructure cost reduction, targeting teams looking to cut cloud embedding API spend without sacrificing index quality.
You can index once with the heavyweight model, then answer live queries with the smaller, faster one.
-
KuCoin News Read →
Provides benchmark scores alongside the compatibility framing, and notes the OCR-bypass and layout-preservation angle for enterprise document pipelines.
The vectors of both models are mutually compatible, allowing developers to build a database using the 9B model while using the 0.6B model for everyday searches
Originally reported by perplexity.ai
Read the original article →Original headline: Perplexity Open-Sources pplx-embed-v2-late, 0.6B and 9B Multimodal Late-Interaction Embedders