pplx-embed-v2-late: Build the Index With the 9B, Query With the 0.6B

Perplexity's new pplx-embed-v2-late models share one embedding space, so you can index documents with the 9B once and serve every live query with the 0.6B at the same query cost as running the small model alone, with better retrieval.

Perplexity released pplx-embed-v2-late last week in two sizes, 0.6B and 9B, under the MIT license on Hugging Face. The detail most teams will miss is that both checkpoints share the same multi vector embedding space. That means an index built with the 9B model can be searched with queries embedded by the 0.6B model, and you do not have to choose one size for the whole pipeline.

Perplexity says this asymmetric setup is the configuration it expects most production deployments to use, and its numbers explain why. On ViDoRe v3 image retrieval, 0.6B queries against a 9B index scored 63.5%, up from 62.3% with the 0.6B model on both sides. Perplexity also says query encoding cost is unchanged between the symmetric 0.6B configuration and the asymmetric one. You pay for the large model once, at indexing time, and every live query costs the same as before.

The second thing to get right is input handling. The models output a 128 dimensional embedding for each token and can search rendered PDF pages without OCR or parsed text, so you can embed page images directly. But the Hugging Face model card says mixed text plus image inputs are not supported in a single encoding call. Keep text only batches and image only batches in separate calls, or your pipeline will fail on the first mixed batch.

Finally, use the right methods. The model card loads the model with the sentence-transformers MultiVectorEncoder and uses encode_query for queries, encode_document for documents, and MaxSim similarity for scoring. Perplexity says the models work with sentence-transformers 6.0.0 or later and transformers 5.4.0 or later, so pin those versions before you start.

Try it today

1. Install sentence-transformers 6.0.0 or later and transformers 5.4.0 or later.
2. Load pplx-embed-v2-late-9b with MultiVectorEncoder and call encode_document over your corpus, sending text pages and page images in separate batches.
3. Store the resulting per token vectors in a multi vector capable index.
4. At query time, load pplx-embed-v2-late-0.6b and call encode_query on each live query.
5. Score with the model's MaxSim similarity against the 9B built index, and compare recall with your current single vector setup on a sample of real queries before switching.