Perplexity AI is putting multimodal retrieval on a smaller hardware budget. On October 7, 2026, it released pplx-embed-v2-late, a pair of ColBERT-style embedding models built to retrieve text, images, and rendered PDF pages from one shared embedding space.
The lineup has a clear split: the 0.6B model handles fast, cheap queries, while the 9B model targets maximum quality. Both models are available on Hugging Face under the MIT license, although a hosted Perplexity API endpoint remains planned rather than live. Cloud convenience can wait its turn.
A Small Query Model With a Large Index
The 0.6B model has 594M total parameters and approximately 240M active parameters. For images, it uses about 340M active parameters and stays close to 8B rivals, while its intended role is a 100% local live query encoder for laptops, edge devices, or small GPUs.
That design supports a two-model retrieval setup. A 9B index can be searched with 0.6B queries, recovering about half of the 9B quality gap on text while keeping the 0.6B query cost. The 9B model has 9B total parameters and approximately 7.4B active parameters, so the division of labor is obvious: smaller model for queries, larger model for indexed quality.
The models also keep their token vectors compact. Instead of compressing an entire document into one vector like dense models, pplx-embed-v2-late stores a 128-dim vector for every token; those vectors are 16x to 32x narrower than rivals using 2,048 to 4,096 dimensions.
Pages are encoded as images, which removes the need for an OCR step. That matters for visual document search across PDFs, slides, and scanned reports, where the page itself carries layout and visual information that plain text extraction can discard.
Strong Benchmarks With Clear Tradeoffs
The 9B model posted the strongest result in the release: 92.4% on MADQA. It also led all tested models by 1.6 percentage points on domain-specific text across 72 tasks using nDCG@10, and both models beat the previous best of 69.3% on Q2D-Web using Recall@1000.
The smaller model still has limits. Its weakest tested result was 61.2% on ViDoRe v3 Markdown, while a 9B index queried by the 0.6B model scored 63.5% on ViDoRe v3 image retrieval. The 0.6B model beat Mixedbread’s retriever at 88.9% but trailed Mixedbread Agentic Search at 93.4%.
Perplexity’s models do not win every comparison. Gemini Embedding 2 beat the 9B model on MIRACL-Vision and led it by 2 percentage points on PPLX-Q2I. That keeps the result in its proper category: a strong retrieval release with specific advantages, not a universal benchmark takeover.
The models were distilled from an 18B teacher using LEAF-style token-level training. Both model cards show CUDA GPU usage, and the published checkpoints are stored in F32, which doubles their download size. Users also need sentence-transformers version 6.0.0 or newer and transformers version 5.4.0 or newer.
Deployment depends on the model selected. The 0.6B version is designed for edge devices and small GPUs, while the 9B version requires a datacenter or high-memory GPU. Both are aimed at low-latency search and agentic RAG over large PDF or web collections, where a lightweight local query encoder can work against a stronger shared index.
The practical pitch is straightforward: use the 0.6B model where query speed, local execution, and hardware limits matter; use the 9B model where retrieval quality justifies the larger footprint. With one embedding space for text, images, and rendered PDF pages, pplx-embed-v2-late gives developers a focused toolkit for visual document search rather than another generic model release wearing a multimodal label.
