Perplexity ships pplx-embed-v2-context-9b-preview

The preview weights are on Hugging Face. The SOTA rankings are Perplexity’s own.

On 30 September 2026 Perplexity announced pplx-embed-v2-context-9b-preview. Each slice of a long document (a chunk) becomes a vector with the whole text in view, in a single pass. That matters for anyone using RAG (retrieval-augmented generation: the system fetches passages and only then answers) on PDFs and contracts, where a slice on its own often loses which clause, year, or annex it came from.

The announcement is in the blog Contextual embedding beyond the gold passage, on @perplexity_ai on X, and in the weights at perplexity-ai/pplx-embed-v2-context-9b-preview on Hugging Face. Perplexity sells search, an API, and embeddings. The context-bench in the post is private to turbopuffer, a vector database named as a partner on the evaluation, not as a co-trainer of the weights.

First: what a contextual embedding is

An embedding turns text into a list of numbers so a computer can find what looks like what. In RAG, a long PDF is sliced up. If the slice is too small, the anchors disappear.

Contextual embeddings try to keep the slice and the surrounding document: a newspaper clipping that still carries the headline. In practice the model reads the document once and averages (pools) the numbers for each chunk. That comparison stops helping when training teaches the model to pull only the gold chunk, the answer span, and treats the rest, including supporting evidence, as a negative.

What pplx-embed-v2-context-9b-preview shipped

According to the blog, at inference time you get one embedding per chunk, with no extra compressor or reranker. Perplexity’s query-aware compressor, described on 14 May 2026, is a teacher only during training. The company starts from an internal ColBERT (a retriever that compares query and document token by token) it calls 9B parameters. The Hugging Face card lists “Model size 8B params.” The exact parameter count is RUMOR.

There is Matryoshka at 1024 and 2048 dimensions (the longer vector contains the shorter one), quantization-aware training for int8, and, in the company’s evaluations, documents of up to 32,768 tokens in one pass. It is a preview, with no compatibility promised, and Perplexity’s API for this model did not ship in this announcement. This is not pplx-embed-v2-late, a multimodal model announced on 7 October 2026.

The numbers, according to Perplexity

The company claims SOTA on ConTEB, Illuin’s public contextual-embedding benchmark, and on turbopuffer’s private context-bench (2,099 queries, 38,894 documents, 2,458,072 chunks, 21 domains). A third-party audited ranking and numbers published by turbopuffer: RUMOR. Perplexity says the run was a blind submission; turbopuffer’s site does not publish the bench.

In the blog’s Figure 1 write-up, at K=10 the preview would have 45.5% answer recall, 40.6% evidence recall and 31.1% all-evidence, and +14.4 percentage points on answer and +5.0 on evidence against Voyage AI’s voyage-context-4 (June 2026). Document@1: 15.2%. Document@10: 61.6%. In the X thread, pplx-embed-context-v1 4B sits slightly ahead on Document@3 and Document@5.

On ConTEB, Perplexity says it has the highest mean nDCG@10 (ranking quality in the first 10 hits) among the models in its chart, without printing the mean in the text and without winning every task. The v1 4B model does better on NarrativeQA. NVIDIA’s Nemotron-3-Embed-8B-BF16, which is not contextual, does better on COVID-QA.

The storage comparison mixes formats: 1 KB per vector (1024 dimensions, int8) against 8 KB for voyage-context-4 (2048 dimensions, float32). The arithmetic 1024 × 1 byte = 1 KB and 2048 × 4 bytes = 8 KB checks out; the “about 8× less” headline compares the smallest int8 vector with the rival’s largest float32 vector.

Limits and what we still do not know

On its internal Q2C/Q2D suites, Perplexity says it leads mean chunk retrieval (legal, technical, health). Voyage-context-4 leads in finance and multilingual; Nemotron-3-8B leads in conversation. On document retrieval the mean sits a little below Voyage. This newsroom has not measured latency, VRAM, or quality outside those benches. MarkTechPost on 30 September 2026 repeats the blog’s figures; that is not an independent measurement.

It fits if you index near-duplicate contracts or long filings and can run the weights yourself. It is not a first-party API in this announcement. A test runs from Hugging Face, with transformers 5.4.0 or newer and trust_remote_code=True.

Context-bench still has no public reproduction. The 9B versus about 8B count is unsettled. There is no API date. What you can do now is download the preview and try it on your own documents, without treating SOTA as an audited fact.

Sources

Official

• Perplexity, Contextual embedding beyond the gold passage, 30 Sep 2026
• @perplexity_ai on X
• Hugging Face, pplx-embed-v2-context-9b-preview
• Perplexity, query-aware compressor, 14 May 2026
• Perplexity, multimodal embeddings beyond a single vector, 7 Oct 2026

Benchmarks and rivals

• ConTEB, Illuin (GitHub)
• turbopuffer
• Voyage AI, Voyage Context 4, 29 Jun 2026
• Hugging Face, NVIDIA Nemotron-3-Embed-8B-BF16
• Hugging Face, pplx-embed-context-v1-4b

Press

• MarkTechPost, 30 Sep 2026