Specialist modelPreview
pplx-embed-v2-context-9b-preview
Perplexity
- Released
- September 30, 2026
- Data date
- October 7, 2026
pplx-embed-v2-context-9b-preview embeds document chunks with the surrounding context of the full document. Each chunk therefore receives a context-aware representation instead of an isolated vector.
Queries and documents use separate encoding methods. The public preview may change its weights, embeddings, and interface without backward compatibility, so its vectors should not be mixed with a later version.
Specifications and access
| Specification | Value and source |
|---|---|
| Model ID | perplexity-ai/pplx-embed-v2-context-9b-previewSource |
| Input modalities | Text queries and text chunks from documents. Use encode for documents and encode_queries for queries.Source |
| Output modalities | One 2,048-dimensional INT8 embedding vector per document chunk; 1,024 or 2,048 dimensions are trainedSource |
| Preview status | Weights, embeddings, and interface may change without backward compatibilitySource |
| License | MIT according to the Hugging Face model cardSource |
| Weight access | Public model weights for local use; MIT licenseSource |
| Hosted service | No Inference Provider on Hugging FaceSource |
| Announcement | Perplexity announced the model on September 30, 2026Source |
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
context-bench Document@1 (Show measurement, test conditions, and source)
- Source value
- 15.2
- Score
- 15.2
- Metric
- Document@1
- Unit
- %
- Benchmark version
- context-bench
- Category
- contextual retrieval
- Direction
- Higher is better
- Sample
- 2099
- Source type
- vendor-reported
- Evaluator
- Perplexity Research and turbopuffer, Contextual embedding beyond the gold passage
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Perplexity preview `pplx-embed-v2-context-9b-preview`; context-bench at K=10 or K=1 for Document@1, with 2,099 queries, 38,894 long documents, and 21 domains. Evaluation encodes documents up to 32,768 tokens in one pass. The post does not say whether these scores used 1,024 or 2,048 dimensions or INT8.
- Context
- Context-bench is privately held by turbopuffer; Perplexity says its preview was evaluated as a blind submission. This remains vendor-reported evidence.
context-bench Document@10 (Show measurement, test conditions, and source)
- Source value
- 61.6
- Score
- 61.6
- Metric
- Document@10
- Unit
- %
- Benchmark version
- context-bench
- Category
- contextual retrieval
- Direction
- Higher is better
- Sample
- 2099
- Source type
- vendor-reported
- Evaluator
- Perplexity Research and turbopuffer, Contextual embedding beyond the gold passage
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Perplexity preview `pplx-embed-v2-context-9b-preview`; context-bench at K=10 or K=1 for Document@1, with 2,099 queries, 38,894 long documents, and 21 domains. Evaluation encodes documents up to 32,768 tokens in one pass. The post does not say whether these scores used 1,024 or 2,048 dimensions or INT8.
- Context
- Context-bench is privately held by turbopuffer; Perplexity says its preview was evaluated as a blind submission. This remains vendor-reported evidence.
context-bench Answer@10 (Show measurement, test conditions, and source)
- Source value
- 45.5
- Score
- 45.5
- Metric
- Answer recall@10
- Unit
- %
- Benchmark version
- context-bench
- Category
- contextual retrieval
- Direction
- Higher is better
- Sample
- 2099
- Source type
- vendor-reported
- Evaluator
- Perplexity Research and turbopuffer, Contextual embedding beyond the gold passage
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Perplexity preview `pplx-embed-v2-context-9b-preview`; context-bench at K=10 or K=1 for Document@1, with 2,099 queries, 38,894 long documents, and 21 domains. Evaluation encodes documents up to 32,768 tokens in one pass. The post does not say whether these scores used 1,024 or 2,048 dimensions or INT8.
- Context
- Context-bench is privately held by turbopuffer; Perplexity says its preview was evaluated as a blind submission. This remains vendor-reported evidence.
context-bench Evidence Recall@10 (Show measurement, test conditions, and source)
- Source value
- 40.6
- Score
- 40.6
- Metric
- Evidence Recall@10
- Unit
- %
- Benchmark version
- context-bench
- Category
- contextual retrieval
- Direction
- Higher is better
- Sample
- 2099
- Source type
- vendor-reported
- Evaluator
- Perplexity Research and turbopuffer, Contextual embedding beyond the gold passage
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Perplexity preview `pplx-embed-v2-context-9b-preview`; context-bench at K=10 or K=1 for Document@1, with 2,099 queries, 38,894 long documents, and 21 domains. Evaluation encodes documents up to 32,768 tokens in one pass. The post does not say whether these scores used 1,024 or 2,048 dimensions or INT8.
- Context
- Context-bench is privately held by turbopuffer; Perplexity says its preview was evaluated as a blind submission. This remains vendor-reported evidence.
context-bench All-Evidence@10 (Show measurement, test conditions, and source)
- Source value
- 31.1
- Score
- 31.1
- Metric
- All-Evidence@10
- Unit
- %
- Benchmark version
- context-bench
- Category
- contextual retrieval
- Direction
- Higher is better
- Sample
- 2099
- Source type
- vendor-reported
- Evaluator
- Perplexity Research and turbopuffer, Contextual embedding beyond the gold passage
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Perplexity preview `pplx-embed-v2-context-9b-preview`; context-bench at K=10 or K=1 for Document@1, with 2,099 queries, 38,894 long documents, and 21 domains. Evaluation encodes documents up to 32,768 tokens in one pass. The post does not say whether these scores used 1,024 or 2,048 dimensions or INT8.
- Context
- Context-bench is privately held by turbopuffer; Perplexity says its preview was evaluated as a blind submission. This remains vendor-reported evidence.
Sources and data date
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 7, 2026. 3 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not found in the inspected sources | ConTEB task-level numerical scores Original wording Perplexity says the preview reaches the highest average on ConTEB, but exact per-task chart values were not available as readable numeric text. |
| Not published | scored embedding dimension and quantization Original wording The post states that the preview supports 1,024 and 2,048 dimensions and INT8, but it does not identify which representation was used for the reported benchmark scores. |
| Additional source | Perplexity pplx-embed model card (retrieved October 7, 2026) · Perplexity pplx-embed model card · Editorial description reviewed October 7, 2026 |
| Additional source | Perplexity API announcement (retrieved October 7, 2026) · Perplexity API announcement · Editorial description reviewed October 7, 2026 |
| Additional source | Perplexity Research and turbopuffer, Contextual embedding beyond the gold passage (retrieved October 7, 2026) |