Voyage Rerank
Voyage AIVoyage AI’s cross-encoder reranking line, currently rerank-2.5 and rerank-2.5-lite: given a query and up to 1,000 candidate documents from a first-stage retriever, they return relevance scores that reorder the shortlist. Both carry a 32,000-token limit on query plus any single document, are multilingual, and accept a natural-language instruction appended to the query to steer relevance.
rerank-2.5 and rerank-2.5-lite are the GA recommendation as of 2026-09-11, but Voyage’s docs now list rerank-3 and rerank-3-lite above them marked “In Preview” and already priced, so the 2.5 line is one step behind the vendor’s own frontier. The Hub repos voyageai/rerank-2.5 and rerank-2.5-lite hold tokenizer files only — no weights — unlike the sibling embedding line, where voyage-4-nano is genuinely open. Verified 2026-09-11 via the Voyage reranker docs, the rerank-2.5 launch post, the pricing page and the Hugging Face model API.
Openness
1 high confidence- weights
- closed(the voyageai/rerank-2.5 Hub repo holds .gitattributes, tokenizer.json and tokenizer_config.json only
- data
- closed(training corpus not released or described)
- code
- closed(no training or inference implementation
- license
- proprietary(hosted rerank endpoint priced per processed token
API-only, so the ladder's first rule fires. Worth recording explicitly that the Hub presence is a decoy: voyageai/rerank-2.5 exists and is ungated, but contains a tokenizer and nothing else. The company that open-weighted voyage-4-nano has open-weighted none of its rerankers.
- https://huggingface.co/api/models/voyageai/rerank-2.5 recorded 2026-09-11
voyageai/rerank-2.5: gated false, 0 downloads, file list is ['.gitattributes', 'tokenizer.json', 'tokenizer_config.json']. No weights, no config, no corpus, no code.
- https://docs.voyageai.com/docs/pricing recorded 2026-09-11
The reranker endpoint is priced per processed token — rerank-2.5 at $0.05 and rerank-2.5-lite at $0.02 per million tokens — which is the only way the models are sold.
Adoption
3 medium confidenceNo artifact to count — the Hub repo records 0 downloads because it holds no weights — and no call volume published. The level rests on distribution: the rerankers ship inside the MongoDB Atlas Embedding and Reranking API alongside the embedding line, on a platform MongoDB says is trusted by tens of thousands of customers, and are separately priced on Voyage's own endpoint.
- https://www.mongodb.com/company/blog/product-release-announcements/introducing-the-embedding-and-reranking-api-on-mongodb-atlas recorded 2026-09-11
MongoDB Atlas's Embedding and Reranking API gives 'direct access to Voyage AI's state-of-the-art retrieval models', including 'high-precision reranking for two-stage retrieval pipelines', on a platform 'trusted by tens of thousands of customers, including over 75% of Fortune 100 companies'.
- https://docs.voyageai.com/docs/pricing recorded 2026-09-11
rerank-2.5 and rerank-2.5-lite carry their own per-token list prices on Voyage's pricing page, so they are sold as a product rather than bundled.
Capability
4 medium confidenceRung 4 rather than 5, on three counts that all point the same way. The 7.94% margin is vendor-run against a vendor-chosen baseline; that baseline is Cohere Rerank v3.5, which Cohere has since replaced with the 4.0 fast/pro pair, so the comparison no longer measures the current rival; and Voyage's own docs now list rerank-3 and rerank-3-lite above the 2.5 line in preview, which is the publisher saying its frontier has moved. Competitive frontier, not the leader. Placed against the rung definition rather than a peer: no single measurement reports this product and a category peer together - a reranker's NDCG@10 against an embedding model's MMTEB - and rule (d) forbids an edge without one.
- https://blog.voyageai.com/2025/08/11/rerank-2-5/ recorded 2026-09-11
'On a standard suite of 93 retrieval datasets spanning multiple domains, rerank-2.5 and rerank-2.5-lite improve retrieval accuracy by 7.94% and 7.16% over Cohere Rerank v3.5.'
- https://docs.voyageai.com/docs/reranker recorded 2026-09-11
Model Choices table: '(In Preview) rerank-3 32,000 Highest accuracy. Recommendation for most applications', '(In Preview) rerank-3-lite', then rerank-2.5 and rerank-2.5-lite at 32,000 tokens as the generalist recommendation; rerank-2 and rerank-1 are listed as legacy.
Verified 2026-09-11