pplx-embed-v2-context-0.6b
by perplexity-ai
Perplexity's small contextual embedding model for text, images and visual documents, embedding each chunk with its document in view.
perplexity-ai/pplx-embed-v2-context-0.6bDeploy pplx-embed-v2-context-0.6b
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
pplx-embed-v2-context-0.6b is a 0.6B-parameter vision-language embedding model from Perplexity for retrieval over text, images and visual documents. In contextual mode it embeds the chunks of a document together, so each chunk's vector reflects the text around it. It outputs 1024-dimensional embeddings and shares an embedding space with the 1024-dimensional prefix of pplx-embed-v2-context-9b, so you can embed documents with one model and query with the other.
Architecture
A fine-tune of pplx-embed-v2-late-0.6b. For images, one forward pass divides the merged visual tokens into whole-row horizontal bands and returns one embedding per band, so a page becomes several searchable chunks that a standard dense ANN index can store. Perplexity's technical report is arXiv 2610.38854.
Mixpeek SDK Integration
# Embed a page image as contextual bands (from the model card).
import torch
from PIL import Image
from transformers import AutoModel
model = AutoModel.from_pretrained(
"perplexity-ai/pplx-embed-v2-context-0.6b",
trust_remote_code=True,
dtype=torch.bfloat16,
attn_implementation="sdpa",
).to("cuda").eval()
with Image.open("page.png") as page:
image_embeddings = model.encode_images(
[page.convert("RGB")],
chunks=8,
min_pixels=16384,
max_pixels=1800000,
normalize_embeddings=True,
)Capabilities
- Contextual and non-contextual text embeddings, 1024 dimensions
- Page images embedded as contextual bands, eight per page in the model card's example
- Separate encode_queries and encode methods for queries and documents
- Shared embedding space with pplx-embed-v2-context-9b
- Int8-quantized output by default
Use Cases on Mixpeek
Performance
We have not measured it. Perplexity publishes benchmark results in its technical report.
Common Pipeline Companions
Frequently Asked Questions
What is a contextual embedding?
An embedding of a chunk computed with the rest of its document in view. A chunk that says 'revenue rose 12%' can then carry which company and year it belongs to. Chunks must not overlap in this mode, per the model card.
How many dimensions does pplx-embed-v2-context-0.6b output?
1024. Compare with cosine similarity, or pass normalize_embeddings=True and use the dot product.
Can it search charts inside PDFs?
It embeds rendered pages as image bands, so a text question can match a page by what it shows. It finds the page. Reading an exact value from the chart takes an OCR or vision model.
What license is it under?
The model card we read on 2026-11-01 states no license. Check with Perplexity before commercial use.
How do I use it with Mixpeek?
Mixpeek stores bring-your-own vectors, so you can embed in your own environment and upsert the 1024-dimensional vectors with the file and page number into a collection.
Specification
Build a pipeline with pplx-embed-v2-context-0.6b
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free