Voyage rerank-2.5
Voyage's instruction-following reranker: 32K context, multilingual, and the default reranker in a growing number of RAG stacks.
About
rerank-2.5 is Voyage AI's generalist reranking model, announced in September 2025 as the successor to rerank-2. It's a cross-encoder that scores query-document pairs, with two headline upgrades: a 32K-token context window (8x Cohere Rerank 3.5, 2x rerank-2) and instruction-following, so you can steer relevance scoring with natural-language instructions. A lighter rerank-2.5-lite trades some accuracy for speed. Voyage reported a 7.94% retrieval-accuracy gain over Cohere Rerank 3.5 across 93 datasets, and 12.7% on the instruction-following MAIR benchmark — vendor numbers, but the model has since become the default reranker in tools like LlamaIndex integrations and the gbrain agent framework.
It's API-only: no open weights. Pricing is $0.05 per million tokens (lite: $0.02), available through the Voyage API and via OpenRouter. Voyage was acquired by MongoDB, and the reranker line has since continued with rerank-3 (September 2026), but 2.5 remains the widely deployed stable pick.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Voyage-reported: +7.94% retrieval accuracy over Cohere Rerank 3.5 across 93 retrieval datasets (provider claim)
- Voyage-reported: +12.70% over Cohere Rerank 3.5 on the Massive Instructed Retrieval Benchmark (MAIR) (provider claim)
- No independent benchmark table verified; third-party projects (e.g. gbrain) adopted it as default based on their own evals
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Voyage announcement Sep 2025 (via MongoDB blog); 32K context + instruction-following corroborated by LlamaIndex integration docs and third-party RAG frameworks that made it their default; pricing $0.05/1M tokens read from OpenRouter model pages Oct 2026. Closed weights. Strengths: real adoption in production RAG stacks, instruction-following is a genuine differentiator, active line (rerank-3 followed in 2026). Docked for: (1) all accuracy claims are Voyage's own with no independent replication found; (2) API-only, no weights, so no self-hosting or offline use. License: proprietary (API terms).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 3.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Announced Sep 2025: 32K context, instruction-following, multilingual (Voyage via MongoDB blog)
- · Default reranker in LlamaIndex VoyageAI integration and the gbrain agent framework
- · Voyage-reported: +7.94% over Cohere Rerank 3.5 across 93 datasets; +12.7% on MAIR (provider claims)
- · API-only, no open weights; $0.05/1M tokens (OpenRouter model pages, Oct 2026)