The launch
On October 7, 2026, Perplexity, known for its AI search engine, published the pplx-embed-v2-late family on Hugging Face: two open embedding models for retrieval, at 0.6B and 9B parameters, under the MIT license (commercial use allowed). A week earlier, the company had released pplx-embed-v2-context.
They are multimodal: they index text and images (photos, screenshots and rendered pages of PDFs and slides) without OCR, preserving tables and layout.
What's different
1. Late interaction (ColBERT-style). Instead of squeezing a whole document into one vector, the model produces one 128-dimensional vector per token or image patch. At search time, each query token looks for the most similar part of the document (MaxSim scoring). Different parts of a query can match different parts of a document, which usually gives more precise results.
2. A shared space across sizes. Both models were distilled token by token from the same internal 18B "teacher." In practice, you can index documents once with the high-quality 9B and run queries with the fast, cheap 0.6B, which can even run on a laptop, without rebuilding the index.
Numbers
- Per Perplexity, on ViDoRe V3 (visual document retrieval), the 0.6B model matches models with five times as many active parameters.
- Per AI Weekly's coverage, the 9B scores 92.4% on MADQA (500 questions over about 18,000 PDF pages).
- The results are the company's own; the full technical report is due "later this year."
Limitations
- An input can be text or an image, not both together.
- Late interaction stores many vectors per document: the index is much larger than with traditional embeddings and needs a vector database that supports multi-vectors.
- The hosted API isn't available yet; for now it's for self-hosters.
Why it matters
Companies keep knowledge in PDFs, slides, scanned spreadsheets and screenshots. OCR pipelines break tables and lose visual context. An open model that searches the page image directly, with a permissive license and a small model for queries, makes enterprise RAG more accurate and cheaper.
How to evaluate it
- Take a real document set (manuals, contracts, reports) and questions with known answers.
- Compare with your current embedding on accuracy and storage cost.
- Test the 9B index + 0.6B query combination.
- Check that your vector database supports multi-vectors (MaxSim).
Sources
- Perplexity blog: https://www.perplexity.ai/en-GB/hub/blog/multimodal-embeddings-beyond-a-single-vector
- Hugging Face, perplexity-ai/pplx-embed-v2-late-9b: https://huggingface.co/perplexity-ai/pplx-embed-v2-late-9b
- AI Weekly (8/10/2026): https://aiweekly.co/alerts/perplexity-open-sources-pplx-embed-v2-late-924-on-madqa
Want better search and RAG over your company's documents? KGA IT offers a free consultation: kga-it.com.