Perplexity 发布 pplx-embed-v2-late 多模态嵌入模型，0.6B 与 9B 两个版本

Perplexity has released pplx-embed-v2-late, a pair of ColBERT-style multimodal embedding models. They come in 2 sizes: 0.6B for fast, cheap queries and 9B for maximum quality. Both models retrieve text, images and rendered PDF pages, and they share one embedding space.

Is it deployable? Yes, if you host it yourself. Both models are on Hugging Face under the MIT license. A hosted Perplexity API endpoint is planned but not live.

TL;DR

The best

The 0.6B model uses about 340M active parameters for images and stays close to 8B rivals; A 9B index can be searched with 0.6B queries, recovering about half the 9B quality gap on text at 0.6B query cost; Its 128-dim token vectors are 16x to 32x narrower than rivals at 2,048 to 4,096 dims; MIT license, with commercial use allowed

The worst

It stores 1 vector per token, so index size grows with document length; It is not #1 on ViDoRe v3 image retrieval; Tencent’s EVIE scores higher; A single input cannot mix text and images; All scores are self-reported, and the technical report is not out yet

Perplexity designed the 0.6B model as a lightweight query encoder that can also run on edge devices. Both model cards show CUDA GPU usage. They need sentence-transformers >= 6.0.0 and transformers >= 5.4.0. The memory figures are our estimate at 2 bytes per parameter, for weights only. The published checkpoints are stored in F32, which doubles the download size.

All numbers below are from Perplexity’s announcement:

The strongest result: 92.4% on MADQA, set by the 9B model. The biggest margin is on BrowseComp+, at 8.7pp over the best dense model.

The weakest result: 61.2% on ViDoRe v3 Markdown, from the 0.6B model. It is still the 2nd-best score on that benchmark. Image search is the real gap: Gemini Embedding 2 beats the 9B model on MIRACL-Vision and by 2pp on PPLX-Q2I.

Mixing sizes: a 9B index queried by the 0.6B model scored 63.5% on ViDoRe v3 image retrieval. That beats 62.3% with 0.6B on both sides, at the same query cost.

How It Works

Dense models compress a document into 1 vector. pplx-embed-v2-late instead keeps a 128-dim vector for every token. It scores with MaxSim: each query token finds its best document token, and those maxima are summed. Pages are encoded as images, so no OCR step is needed. Perplexity distilled both models from an 18B teacher using LEAF-style token-level training. That training is what creates the shared space.

Best Use Cases

The best fit is visual document search over PDFs, slides and scanned reports; Low-latency search: index in the cloud with 9B, then query on-device with 0.6B; Agentic RAG over large PDF or web collections

The 0.6B model fits edge devices; the 9B model is built for index-time quality; Best score: 92.4% on MADQA. Weakest: 61.2% on ViDoRe v3 Markdown; 0.6B queries over a 9B index beat 0.6B on both sides; Storage growth and self-reported scores are the main caveats

[Sponsored] The web is the one API most agents are missing. Databases, calendars and repos have APIs. The open web mostly doesn’t. The TinyFish MCP server gives any MCP client four tools: TinySearch, TinyFetch (full pages as markdown, JavaScript included), TinyBrowser for logins and forms, and TinyAgent for multi-step jobs. Search and Fetch are free.

Asif Razzaq

Asif RazzaqWhat Happens When a Trusted Model Repo Changes? Unsloth Studio Re-Checks Before It Runs; Asif RazzaqLiquid AI Releases Open-Weight d1-3B and d1-omni-600M: Multimodal Decision Models With Zero Output Tokens; Asif RazzaqMeta AI Open-Sources Rebalancer: A C++ Assignment Solver That Runs About 40 Million Placement Problems a Day; Asif RazzaqGoogle DeepMind Releases EmbeddingGemma 2, a 740M Open Multimodal Embedding Model Built on Gemma 4