Specialist modelActive
Voyage Rerank 3
Voyage AI
- Released
- September 1, 2026
- Data date
- October 7, 2026
Voyage Rerank 3 takes search queries with previously retrieved text documents and sorts the candidates by relevance. Compared with Rerank 2.5, Voyage reports an updated backbone and an improved mix of training data.
The active release began on September 29 after a preview from September 1, 2026. Voyage says the API behavior, 32,000-token context limit, and price stayed the same as Rerank 2.5.
Specifications and access
| Specification | Value and source |
|---|---|
| API model ID | rerank-3Source |
| Lifecycle | Preview from September 1, active from September 29, 2026Source |
| Input modalities | Text query and list of text documentsSource |
| Output modalities | Documents sorted by descending relevance scores, not embedding vectorsSource |
| Context window | 32,000 tokensSource |
| Focus | Voyage's highest-accuracy reranker, optimized for long documents and code retrievalSource |
Page 1 of 22
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.
Voyage Rerank 3 pipeline, 9-domain average, first stage voyage-3-large (Show measurement, test conditions, and source)
- Source value
- 86.9
- Score
- 86.9
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- 9-domain average retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-3-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, Technical documentation, first stage voyage-3-large (Show measurement, test conditions, and source)
- Source value
- 90.8
- Score
- 90.8
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- Technical documentation retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-3-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, Web reviews, first stage voyage-3-large (Show measurement, test conditions, and source)
- Source value
- 76.86
- Score
- 76.86
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- Web reviews retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-3-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, Code, first stage voyage-3-large (Show measurement, test conditions, and source)
- Source value
- 85.48
- Score
- 85.48
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- Code retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-3-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, Medical, first stage voyage-3-large (Show measurement, test conditions, and source)
- Source value
- 87.71
- Score
- 87.71
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- Medical retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-3-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, Conversation, first stage voyage-3-large (Show measurement, test conditions, and source)
- Source value
- 94.3
- Score
- 94.3
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- Conversation retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-3-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, Law, first stage voyage-3-large (Show measurement, test conditions, and source)
- Source value
- 77.97
- Score
- 77.97
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- Law retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-3-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, Finance, first stage voyage-3-large (Show measurement, test conditions, and source)
- Source value
- 89.57
- Score
- 89.57
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- Finance retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-3-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, Long documents, first stage voyage-3-large (Show measurement, test conditions, and source)
- Source value
- 92.49
- Score
- 92.49
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- Long documents retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-3-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, 9-domain average, first stage voyage-4-large (Show measurement, test conditions, and source)
- Source value
- 87.05
- Score
- 87.05
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- 9-domain average retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-4-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, Technical documentation, first stage voyage-4-large (Show measurement, test conditions, and source)
- Source value
- 90.81
- Score
- 90.81
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- Technical documentation retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-4-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Voyage Rerank 3 pipeline, Web reviews, first stage voyage-4-large (Show measurement, test conditions, and source)
- Source value
- 77.6
- Score
- 77.6
- Metric
- NDCG@10
- Unit
- %
- Benchmark version
- Voyage standard retrieval suite, 95 datasets
- Category
- Web reviews retrieval
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Voyage AI public Rerank 3 evaluation spreadsheet
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Pipeline result with first-stage search voyage-4-large and Voyage reranker rerank-3. Up to 100 candidate documents per query were reranked to the top 10; NDCG@10. The suite contains 95 datasets across nine domains. This is not a standalone reranker score.
- Context
- The publisher measures the two-stage pipeline. The result depends on the named first-stage candidate retriever and is not a standalone reranker score.
Page 1 of 22
Published prices
- Processed-token price
- $0.05 per 1M processed tokens; billing counts query tokens per document plus all document tokens. The first 200M tokens are free.Source