Perplexity just released two free search models, and pplx-embed-v2-late can read charts, tables and PDF pages without any text-extraction step. The open weights landed on Hugging Face on October 7, 2026, in a 0.6B and a 9B size.
If you build search or “chat with your documents” tools, this is worth a look. If you just use apps, it explains why document search may get better soon.
Here are the facts you need first:
- Announced October 7, 2026, in two sizes: 0.6B and 9B parameters.
- Handles text, images, tables and visual documents such as PDFs and slide decks.
- Open weights on Hugging Face, listed as MIT-licensed with commercial use allowed.
- Covers 46 languages, according to a Crypto Briefing report.
Those four lines cover most of what a developer needs to decide whether to keep reading.
What pplx-embed-v2-late actually does
An embedding model turns text or images into numbers so a computer can find similar items fast. Search tools use these numbers to match your question with the right document.
Most older setups first convert a PDF to plain text, then embed that text. Charts and tables often get mangled in that step. This model skips it and reads the page as an image, so a bar chart or a pricing table keeps its meaning.
Perplexity says the models were distilled from an 18B teacher model and trained on 186 million query-document pairs. That detail comes from the same Crypto Briefing report.
Why “late interaction” matters for document search
Standard embedding models squeeze a whole page into one vector. That is cheap, but it loses detail. Late interaction keeps one small vector for every token, then compares your query to each of them.
This ColBERT-style method is called MaxSim scoring. Each vector here has 128 dimensions, per the LLM Reference model page. The payoff is finer matching, so a question about one cell in a table can still find the right page.
The cost is storage. Per-token vectors take far more space than one vector per page, and that bill grows with your corpus. It is the main trade-off to weigh before you switch.
0.6B or 9B: which size should you pick?
The two checkpoints share one embedding space. That means you can index your files with the big model and search them with the small one. The table below shows how the sizes compare.
| Detail | 0.6B model | 9B model |
|---|---|---|
| Best for | Edge devices, fast query-time search | Indexing large archives with top accuracy |
| Embedding space | Shared with the 9B model | Shared with the 0.6B model |
| Vector size | 128 dimensions per token | 128 dimensions per token |
| License | MIT, open weights | MIT, open weights |
The mix-and-match trick is the interesting part. You pay the heavy compute once at indexing time, then keep daily searches light. This matches the open-model trend we covered in our Cloudflare Clef explainer.
What the benchmarks say, and what they don’t
Perplexity reports that visual document retrieval on ViDoRe v3 rose from 62.3% to 63.5%. It also reports the 9B model reaching up to 81.3% average nDCG@10 on some retrieval tasks, per Crypto Briefing.
That 1.2-point gain is real but small. Treat these as vendor numbers until independent labs repeat them. Also note that mixed text-plus-image batches are not supported, according to the LLM Reference page.
There is one more catch. Perplexity’s blog says a hosted API is coming, but no pricing is public yet. For now you must run the weights yourself, and our guide to the LMCache vulnerability in vLLM servers is a reminder to keep self-hosted AI infrastructure patched.
Who should try it, and how to test it safely
The best fit is any team with a pile of slide decks, scanned reports or legal PDFs. In those files, the meaning often sits inside a chart or table rather than the body text, which is exactly where older pipelines fail.
A sensible test takes an afternoon. Pick 50 real pages, write 20 questions you know the answers to, and compare results against your current search setup. Check hit quality and index size side by side.
Perplexity’s own numbers look strong on some tests. According to AI Weekly, the pair scores 92.4% on MADQA, a document question-answering test. That is a vendor-reported result, so your own data matters more than any leaderboard. For another open-weight release with a similar “check the license first” lesson, see our Mistral Large 4 open-weights post.
Frequently Asked Questions
Is pplx-embed-v2-late free to use?
The weights are free to download on Hugging Face, and the model page lists an MIT license with commercial use allowed. You still pay for the hardware you run it on.
Can it search PDFs and slide decks?
Yes. It reads visual documents directly, so you don’t need OCR or text parsing first. Mixed text and image batches are not supported.
Can I search a 9B index with the 0.6B model?
Yes. Both models share one embedding space, so the small model can query an index built by the large one.
Is there a hosted API?
Not yet. Perplexity says one is coming, but pricing hasn’t been published.
Our take
If your documents are full of charts and scanned tables, test the 0.6B model on a small sample this week. Skip the full 9B rollout until you’ve measured storage costs, because per-token vectors add up quickly. For everyone else, the news is simply that searching inside PDFs should get less frustrating.
Leave a Reply