Voyage AI releases rerank-3 and rerank-3-lite with gains on long documents and code

What the new models do

Rerankers are the second stage of a retrieval pipeline. The first stage, usually embeddings and vector search, picks candidates from a large corpus. The reranker then scores each candidate against the query and reorders the list so the most relevant documents come first. Voyage AI says that because the reranker sees only a short list, it can be more accurate per document than an embedding model, and it can be added to an existing system with one API call and no re-indexing.

The company says rerank-2.5, released in August 2025, has been one of its most popular models. Since then, coding agents have come to issue many of the queries reaching its rerankers, and the documents being reranked are getting longer. Rerank 3 targets those two areas. Both models use an updated backbone and an improved mixture of training data, according to Voyage AI.

The series is meant as a drop-in upgrade. The API, the 32K token context length, instruction-following and pricing stay the same. Relevance scores are calibrated to match the score distributions of the matching Rerank 2.5 models, so score thresholds tuned on Rerank 2.5 should keep working.

Benchmark results reported by Voyage AI

On the company's suite of 95 retrieval datasets across 9 domains, averaged across four first-stage retrieval methods, rerank-3 outperforms Cohere Rerank v4.0 Pro by 2.72%, Cohere Rerank v4.0 Fast by 5.74%, Qwen3-Reranker-8B by 3.02% and rerank-2.5 by 0.96%. Rerank-3-lite beats Cohere Rerank v4.0 Pro by 2.23%, Cohere Rerank v4.0 Fast by 5.25%, Qwen3-Reranker-8B by 2.53% and rerank-2.5-lite by 1.10%. Voyage AI notes that rerank-3-lite outperforms Qwen3-Reranker-8B, which it calls the leading open-weights reranker, despite being over an order of magnitude smaller.

• Long documents: On LongEmbed (NarrativeQA, SummScreenFD and QMSum), rerank-3 beats rerank-2.5 by 3.35% and Cohere Rerank v4.0 Pro by 13.84%. Rerank-3-lite beats rerank-2.5-lite by 1.86%. The gains over rerank-2.5 come from NarrativeQA and QMSum, where rerank-3 improves by 3.77% and 6.20%.
• Code: Rerank-3 outperforms rerank-2.5 by 2.17% atop voyage-3-large and 2.11% atop voyage-4-large. Rerank-3-lite beats rerank-2.5-lite by 3.72% and 3.58%.
• Multilingual: On 51 datasets, rerank-3 leads rerank-2.5 by 0.52% and Cohere Rerank v4.0 Pro by 1.19%.
• Instruction-following: On MAIR, rerank-3 is on par with rerank-2.5, and rerank-3-lite beats rerank-2.5-lite by 0.91%.

Voyage AI also reports that rerank-3 improves NDCG@10 by 7.65% atop voyage-3-large, 6.38% atop voyage-4-large, 16.67% atop OpenAI v3 large and 20.33% atop BM25. In a comparison on 26 datasets, it says rerank-3 leads the Jev 1.13 model by 1.77, 5.69 and 5.87 NDCG@10 points in code, law and finance. The company adds that it did not run a full evaluation of Jev because of rate limits.

The evaluation retrieves up to 100 candidates per query with a first-stage method, reranks them and keeps the top 10, measured with NDCG@10. All figures come from the company's own tests.

Pricing, availability and what it means for users

Rerank-3 and rerank-3-lite are available now at the same prices as their predecessors: $0.05 and $0.02 per million tokens. Voyage AI says rerank-3-lite now roughly matches the retrieval quality of rerank-2.5 at 40% of the price. New users get the first 200M tokens free. The models are also offered to MongoDB Atlas customers through the Atlas Embedding and Reranking API and Native Reranking in Atlas Vector Search.

For teams already running rerank-2.5 or rerank-2.5-lite, Voyage AI recommends moving to the matching Rerank 3 model, saying the upgrade brings better quality at the same cost with no code changes. The company will keep offering the rerank-2.5 series for existing users who prefer not to switch. Teams with long documents or code-heavy search workloads, the areas where the company reports the biggest improvements, are the most obvious candidates to test the new models.