Perplexity AI Updates: October 1, 2026

1. Perplexity Releases pplx-embed-v2-context-9b-preview, an Open Contextual Embedding Model

Perplexity. Perplexity Research and turbopuffer released preview weights for pplx-embed-v2-context-9b-preview under the MIT license, a contextual embedding model that uses late chunking (encoding a whole document before pooling per chunk) and is trained to retrieve both an answer and the evidence needed to verify it rather than a single gold passage. It outputs 2048- or 1024-dimension Matryoshka vectors, and on turbopuffer’s context-bench at K=10 it reports 45.5% answer recall and 40.6% evidence recall, 14.4 points ahead of voyage-context-4 on answer recall, with 1 KB int8 vectors slightly beating Voyage’s 8 KB float32 vectors. The model is self-hosted only for now (transformers 5.4.0 or later with trust_remote_code), and Perplexity API support has not shipped yet. Source