A team of ten authors posted a technical report on arXiv on 21 October 2026 describing pplx-embed-v2-late. The work adds two new models – a 0.6 billion‑parameter version and a 9 billion‑parameter version – that can handle text, images and scanned documents in many languages. They claim the models beat all other systems of comparable size on several retrieval tasks, and the smaller model runs fast enough for real‑time queries.
How the models work together
Both models were trained by “representation distillation” from a larger 18 billion‑parameter teacher. That means the big teacher taught the two new models to produce similar embeddings – the numeric vectors used for searching – while being much smaller. Because the embeddings share the same space, users can index a collection with the 9 B model (which creates more accurate vectors) and later search it with the lightweight 0.6 B model, which responds quickly.
Where to get them
The authors provide download links in the report’s “this https URL” line. No pricing or licensing details are given in the paper, so the models appear to be released for free academic or research use.
Why it matters
For developers building search engines or recommendation tools, the pair lets them keep indexing quality high with the 9 B model while serving user queries fast with the 0.6 B model. This could lower compute costs for multilingual, image‑rich applications, though independent benchmarks are not yet available.
