pplx-embed-v2-late-0.6b

pplx-embed-v2-late-0.6b

Scores queries against text, images and document pages with a ColBERT-style model that keeps one small vector per token. For developers who build search over mixed documents.

Made by perplexity-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: pplx-embed-v2-late-0.6b is the smaller model in perplexity-ai's pplx-embed-v2-late family of multimodal retrievers. It is built on Qwen3.5 and gives each token its own 128-dimension vector. A query and a document are then matched with MaxSim, a late-interaction score. It shares an embedding space with the 9B model, so this one can query an index built by the larger one.

Who it suits: Developers who build search over text, images and document pages, and who want token-level matching with a small model.

Industry | Who uses it | How they use it | Example
Software | Search engineers | Rank passages for a query with late interaction | Score a short question against a list of text passages
Document-heavy teams | Developers with scanned pages | Embed page images and match them to text queries | Find the page that answers a written question
Legal research | Search tool builders | Match legal questions to source passages | The source's own query asks what statute governs limitations

Anyone who needs a chat model or text generation can skip it, because it only produces vectors and scores.

Where to use it:

• Python with Sentence Transformers: the model loads through the MultiVectorEncoder class.
• A CUDA GPU: the source's example sets the device to cuda.
• An index built with the 9B model: the 0.6B model can query it.
• Windows, macOS or Linux: the download steps run on all three with Python.

Use case | Who | What they get back
Text retrieval | Search developers | MaxSim scores between each query and each passage
Page image retrieval | Developers with scanned pages | MaxSim scores between a text query and each page image
Cheap queries on a big index | Teams using the 9B model for indexing | A smaller model that can answer queries against that index

How to set it up: What you need first:

• Python with pip.
• sentence-transformers 6.0.0 or newer.
• transformers 5.4.0 or newer.
• Pillow, if you embed images.

1. Install the two packages with pip.
2. Optionally download the model files with the hf command.
3. Load the model with MultiVectorEncoder, using the 0.6B or the 9B name.

pip install 'sentence-transformers>=6.0.0' 'transformers>=5.4.0'

pip install -U huggingface_hub
hf download perplexity-ai/pplx-embed-v2-late-0.6b

How to use it: Encode queries with encode_query and documents with encode_document, then call similarity to get MaxSim scores. Send text-only and image-only batches in separate calls, because a batch that mixes both is not supported. The export uses native Sentence Transformers modules, so it needs no custom Python code.

from PIL import Image
from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder(
    "perplexity-ai/pplx-embed-v2-late-0.6b",  # or "perplexity-ai/pplx-embed-v2-late-9b"
    device="cuda",
)
queries = model.encode_query(["what statute governs limitations?"])
documents = model.encode_document(["A document passage"])
scores = model.similarity(queries, documents)  # MaxSim

image = Image.open("document.png").convert("RGB")
image_documents = model.encode_document([image])
image_scores = model.similarity(queries, image_documents)

Title: Calls the source shows
Tool | What it does
MultiVectorEncoder | Loads the model with its multi-vector settings
encode_query | Turns queries into one vector per token
encode_document | Does the same for text passages or images
similarity | Scores queries against documents with MaxSim

Example: You ask "what statute governs limitations?" and compare it with a text passage and then with a scanned page opened as an image. Each comparison returns a MaxSim score. You keep text and image documents in separate encoding calls.

For newcomers and experts: Most embedding models squeeze a whole text into one vector. This one keeps a small vector for every token, then lines up each query token with its best match in the document. A good first try is to run the sample code with one query and one short passage and read the score.

An internal ColBERT teacher with 18B parameters trained both sizes by distillation, with a token-level objective in the style of LEAF. In the 0.6B model, every weight was updated. In the 9B model, only the last eight transformer layers were updated in full and the rest used LoRA. PyLate puts query and document markers second, but this model expects them first.

Encode the query into one vector per token
Encode each document into one vector per token
Match every query token to its best document token
Sum the matches into a MaxSim score

Good to know: The card compares the two sizes on a public page-retrieval benchmark.

Title: ViDoRe v3 nDCG@10 on page images
0.6B model: 62.3 percent
9B model: 65.2 percent

• Active size: the card lists 340M active parameters for this model and 7.4B for the 9B one.
• Markdown input: the same benchmark gives 61.2 percent for the 0.6B and 64.7 percent for the 9B.
• Inputs: text and images are separate, and mixed batches are not supported.
• License: mit
• Cost to run: no API key, account or paid service named.
• Data it reaches: files you download from Hugging Face, and the text and images you encode.
• Not for: text generation, or batches that mix text and images.

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: sentence-transformers
• Parameters: 594M
• Weights: about 2.4 GB in F32; about 297 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,765
• Likes: 71
• License: mit
• Quantized versions: 4
• Fine-tunes: 6
• Tags: sentence-transformers, safetensors, qwen3_5, feature-extraction, sentence-similarity, mteb, multilingual

Links

• Model: https://huggingface.co/perplexity-ai/pplx-embed-v2-late-0.6b
• Files: https://huggingface.co/perplexity-ai/pplx-embed-v2-late-0.6b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download perplexity-ai/pplx-embed-v2-late-0.6b

Then follow the usage in its model card. Nothing in this kit runs by itself.