Is pplx-embed-v2-context state of the art?Pending benchmarks.Announced Oct 28, 2026; no independent results yet.
Perplexity's 0.6B and 9B multilingual vision-language contextual embedding models: each chunk of a document is embedded with the rest of the document in view, as one int8 vector per chunk or per document. Images can be split into horizontal bands and embedded as chunks. The 0.6B shares an embedding space with the 9B's 1024-dimensional prefix. Developed with turbopuffer; open weights on Hugging Face.
The details so far
- Lab
- Perplexity
- Announced
- Oct 28, 2026
- API
- Not yet · None announced by Perplexity
- Open weights
- Published · Hugging Face perplexity-ai/pplx-embed-v2-context-0.6b and -9b, not gated (F32 safetensors, custom code); needs transformers 5.4+
- Parameters
- 8,401,083,632 (9B) and 595,370,176 (0.6B) stored parameters
- Sizes
- pplx-embed-v2-context-9b and pplx-embed-v2-context-0.6b
- Base model
- Fine-tuned from pplx-embed-v2-late-9b and pplx-embed-v2-late-0.6b (Hugging Face card metadata)
- Type
- Dense contextual embeddings: one vector per chunk, with the surrounding document in view; a non-contextual mode embeds text on its own
- Vectors
- Unnormalized int8; 2048 dimensions for the 9B with Matryoshka prefixes at 1024 and 2048; 1024 dimensions for the 0.6B
- Inputs
- Text, images and visual documents; images can be split into up to 16 horizontal bands, each embedded as a chunk
- Languages
- Multilingual
- Shared embedding space
- The 0.6B matches the 9B's 1024-dimensional prefix, so one can index and the other can query
- Training
- 9B: chunk-level distillation of relevance scores from Perplexity's query-aware context compression model plus a document-level contrastive loss, trained on text only; 0.6B: distilled from the 9B; quantization-aware training for int8
- License
- MIT on the 9B card; the 0.6B card states no license. The paper is CC BY-NC-ND 4.0
- Paper
- Contextual Embeddings Beyond the Gold Chunk (arXiv 2610.38854), Perplexity AI with William Barber of turbopuffer
- New benchmarks
- context-bench (run blind by turbopuffer) and Q2D-Chunk (public leaderboard on Hugging Face)
- Preview
- pplx-embed-v2-context-9b-preview, a separate earlier checkpoint, was announced on Perplexity's blog on Sep 30, 2026
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: on context-bench, evaluated blind by turbopuffer, the 9B is 14.4 points ahead of voyage-context-4 at Answer@10.
Lab claim: on ViDoRe-V3, with two chunks per image and 1024-dimensional vectors, the 9B outperforms Nemotron-3.5-8B with a 4x smaller index.
Lab claim: the 9B scores highest of the models Perplexity compared on context-bench, Q2D-Chunk and ConTEB.
Sources
PAPERContextual Embeddings Beyond the Gold Chunk (arXiv)arxiv.org/abs/2610.38854 · Oct 28, 2026REPOperplexity-ai/pplx-embed-v2-context-0.6b on Hugging Facehuggingface.co/perplexity-ai/pplx-embed-v2-context-0.6bREPOperplexity-ai/pplx-embed-v2-context-9b on Hugging Facehuggingface.co/perplexity-ai/pplx-embed-v2-context-9bLABPerplexity Developers on Xx.com/perplexitydevs/status/2117373903287050675 · Nov 2, 2026REPOQ2D-Chunk leaderboard (Hugging Face Space)huggingface.co/spaces/perplexity-ai/q2d-chunk-leaderboardTracking continues until an independent board lists it.