Best open-source reranker models in 2026

A reranker reads a question together with each candidate result and re-scores a short list, which is slower than an embedding search but more accurate. The usual recipe is to fetch 50 to 100 candidates with an embedding model and let a reranker sort them. Most rerankers here are small enough for one GPU, and the license decides whether you can ship one.

  • List updated November 11, 2026
  1. Qwen3-Reranker Best overall

    Alibaba's Apache 2.0 rerankers in 0.6B, 4B and 8B sizes, with 100+ languages, code, a 32K context and custom instructions. In Qwen's own table the 8B leads four of six columns by small margins and the 4B beats it on the other two, and the 4B is the most downloaded of the three.

    • Alibaba Qwen
    • 8.19B
    • 32K tokens
    • Apache 2.0 (Permissive)
    • from 1.37 GB
  2. BGE Reranker v2 m3 Most downloaded

    BAAI's 568M multilingual cross-encoder, with about 14.7 million downloads in 30 days, far more than any other reranker here. Apache 2.0, built on BGE-M3, and in Qwen's June 2025 table a newer 0.6B reranker is ahead of it except on Chinese retrieval.

    • BAAI
    • 568M
    • 8K tokens
    • Apache 2.0 (Permissive)
    • from 2.61 GB
  3. Jina Reranker v3.5

    Jina's 0.6B listwise reranker that scores a whole candidate list in one pass, with 131K context and a 0.40 GB Q4_K_M GGUF. It targets legal, medical and structured data, but the license is CC BY-NC and Qwen3-Reranker-4B is ahead on three of four columns in Jina's own table.

    • Jina AI
    • 597M
    • 128K tokens
    • CC BY-NC (Non-commercial)
    • from 0.46 GB
  4. Qwen3-VL-Reranker

    Alibaba's Apache 2.0 2B and 8B rerankers for queries and candidates that can be text, images, screenshots or video. In Qwen's own table the 8B scores 86.3 on visual documents, and the 2B trails jina-reranker-m0 on two visual-document tests.

    • Alibaba Qwen
    • 8.77B
    • 32K tokens
    • Apache 2.0 (Permissive)
    • from 4.9 GB
  5. Nemotron 3.5 Rerank 8B Newest

    NVIDIA's 8B reranker for text and document-page images under the permissive OpenMDW license, released on 20 October 2026. NVIDIA reports 60.49 on the ViDoRe v3.1 reranking leaderboard against 59.17 for Qwen3-VL-Reranker-8B, and the BF16 file is 16.76 GB.

    • NVIDIA
    • 8.38B
    • 32K tokens
    • OpenMDW-1.1 (Permissive)
    • from 19 GB

How we chose

  • Open weights you can download today from the developer's official repository.
  • Languages, context length and the kinds of input are listed, because they decide whether a reranker fits your documents.
  • Scores come from the developers and use different candidate lists and rivals, so test two or three rerankers on a few hundred real queries from your own data.
  • We read each license, because one reranker here is free for non-commercial use only.