Perplexity releases pplx-embed-v2-context, open-weight vision-language embedding models in 9B and 0.6B

Perplexity released pplx-embed-v2-context, two open-weight vision-language embedding models (9B and 0.6B) on Hugging Face, with a paper written with turbopuffer. Both return one int8 vector per document, or one per chunk that sees the whole document, and both also run in a non-contextual mode with one embedding per document and no chunking.

Perplexity says it trained them differently from other contextual embedding models: chunk-level relevance scores are distilled from a query-aware context compression model in its answer pipeline, instead of one gold chunk per query, so supporting chunks are no longer treated as negatives.

Results below are Perplexity's own:

• The 9B leads ConTEB and two new benchmarks, context-bench and Q2D-Chunk. Q2D-Chunk is built from real Perplexity search traffic (6,069 queries, 10 languages, 71,100 documents) and labels both answer chunks and the evidence needed to verify them. Perplexity published a leaderboard for it.
• On ViDoRe-V3 markdown, the 0.6B is the strongest sub-1B model, chunked or unchunked. The contextual 9B with 128-token chunks is second only to voyage-context-4.
• The 9B was trained only on text, yet Perplexity says it outperforms the leading dense retrievers on visual documents. With two 1024-dimension embeddings per page it beats Nemotron-3.5-Embed-8B on ViDoRe-V3 with an index four times smaller.