Los mejores modelos de reclasificación para búsqueda y RAG
Clasificaciones de modelos actualizadas según datos de uso reales.
Los modelos se clasifican según el total de tókenes de prompts y respuestas procesados a través de la API de OpenRouter durante los últimos 7 días. Se muestran los 10 modelos mejor clasificados. Las clasificaciones miden el uso en OpenRouter y reflejan la adopción, no la calidad del modelo ni su rendimiento en pruebas de evaluación.
Los modelos de reclasificación reordenan documentos candidatos, pasajes o resultados de búsqueda según su relevancia, lo que mejora la etapa de recuperación en la búsqueda semántica, la generación aumentada por recuperación (RAG) y los flujos de recomendación. Esta colección clasifica los modelos de reclasificación según su uso en OpenRouter durante la última semana. Los mejores modelos actuales son rerank-2.5-lite, rerank-2.5 y rerank-3-lite. Compare precios, latencia y calidad para encontrar la mejor opción para su flujo de búsqueda o RAG.
Principales modelos de reclasificación en OpenRouter
VoyageAI by MongoDB: rerank-2.5-lite
65,2B tókenes
rerank-2.5-lite is a reranker optimized for both latency and quality, delivering a 7.16% improvement in retrieval accuracy over Cohere Rerank v3.5 across 93 datasets. It also outperformed Cohere Rerank v3.5 by 10.36% on the Massive Instructed Retrieval Benchmark (MAIR). The model supports a combined context length of 32K tokens per query–document pair, including up to 8K tokens for the query, enabling more accurate retrieval over longer documents. Additionally, rerank-2.5-lite supports instruction following, allowrerank-2.5-lite here: mongodb.com/products/platform/ai-search-and-retrieval/models
VoyageAI by MongoDB: rerank-2.5
32,1B tókenes
rerank-2.5 is a cutting-edge reranker optimized for quality, delivering a 7.94% improvement in retrieval accuracy over Cohere Rerank v3.5 across 93 datasets. It also outperformed Cohere Rerank v3.5 by 12.70% on the Massive Instructed Retrieval Benchmark (MAIR). The model supports a combined context length of 32K tokens per query–document pair, including up to 8K tokens for the query, enabling more accurate retrieval over longer documents. Additionally, rerank-2.5 supports instruction following, allowingbout rerank-2.5 here: mongodb.com/products/platform/ai-search-and-retrieval/models
VoyageAI by MongoDB: rerank-3-lite
17B tókenes
rerank-3-lite is a reranker optimized for both latency and quality and a drop-in upgrade to rerank-2.5-lite, improving on it by 0.94% NDCG@10 on average across domain evaluations and by 1.86% on long-document evaluations, with code retrieval gains of 2.77% atop voyage-3-large and 2.59% atop voyage-4-large. It matches the retrieval quality of rerank-2.5, and across 93 retrieval datasets it outperforms Cohere Rerank v4.0 Pro by 1.44% and Qwen3-Reranker-8B by 2.61%. The model supports a combined context length of 32K tokens per query-document pair, including up to 8K tokens for the query, enabling more accurate retrieval over longer documents. Additionally, rerank-3-lite supports instruction following, allowing users to guide relevance scoring t rerank-3-lite here: mongodb.com/products/platform/ai-search-and-retrieval/models
Qwen3 Reranker 8B
9,4B tókenes
Qwen3 Reranker 8B is a text reranking model from Alibaba Cloud built on the Qwen3 architecture. It evaluates query-document pairs to produce relevance scores for use in retrieval and RAG pipelines. Supports 100+ languages and programming languages, with instruction-aware reranking that allows customizing scoring criteria per task. Offers strong performance on multilingual benchmarks including MTEB, CMTEB, and MMTEB.
VoyageAI by MongoDB: rerank-3
8,62B tókenes
rerank-3 is a reranker optimized for quality and a drop-in upgrade to rerank-2.5, improving on it by 0.80% NDCG@10 on average across domain evaluations and by 3.35% on long-document evaluations, with code retrieval gains of 2.01% atop voyage-3-large and 1.96% atop voyage-4-large. Across 93 retrieval datasets it outperforms Cohere Rerank v4.0 Pro by 2.14% and Qwen3-Reranker-8B by 3.31%. The model supports a combined context length of 32K tokens per query-document pair, including up to 8K tokens for the query, enabling more accurate retrieval over longer documents. Additionally, rerank-3 supports instruction following, allowing users to guide relevance sco about rerank-3 here: mongodb.com/products/platform/ai-search-and-retrieval/models
NVIDIA: Llama Nemotron Rerank VL 1B V2 (free)
1,63B tókenes
Llama Nemotron Rerank VL 1B V2 is a 1.7B multimodal reranking model from NVIDIA. It evaluates the relevance of document images and text against user queries, designed for vision RAG pipelines handling charts, tables, infographics, and mixed-media documents. Functions as a cross-encoder that accepts text queries paired with image, text, or combined document inputs, delivering approximately 6-7% recall improvements over embedding-only baselines on visual document retrieval benchmarks.
NVIDIA: Nemotron 3.5 Rerank 8B (free)
989M tókenes
Cohere's AI search foundation model for enhancing the relevance of information surfaced within search and RAG systems. Features a 32K context window, multilingual support across 100+ languages, no data pre-processing required, and state of the art performance with low latency.
Cohere's AI search foundation model for enhancing the relevance of information surfaced within search and RAG systems. Features a 32K context window, multilingual support across 100+ languages, no data pre-processing required, and high performance with lowest latency.
Rerank v3.5 is designed to reorder search results for improved relevance. It supports multi-aspect and semi-structured data reranking over 100+ languages. Ideal for refining results from semantic or keyword search pipelines.
Preguntas frecuentes
Send POST /api/v1/rerank with the model slug, your query, a documents array, and optionally top_n. The response returns results with a relevance_score and index for each document. The RAG cookbook walks through the request.
Listings with published pricing bill per search unit or per input token, depending on the model, so compare prices in the same unit before you choose. You pay only for the requests you make, with no separate subscription per provider.
Retrieve candidate chunks with an embedding model, rerank the top candidates to sharpen precision, then pass the best documents to an LLM as context. All three steps run through the same OpenRouter API key.
Yes. One OpenRouter API key works for every rerank model on this page and for every other model on OpenRouter. Create a key on your API keys page.