Qwen3 Reranker

weights

open(Apache-2.0 ungated safetensors across the three text cross-encoder checkpoints)

data

closed(no training corpus or data-construction scripts released)

code

partial(the Qwen3-Embedding repository's docs/training directory holds a single file, SWIFT.md, which documents fine-tuning through the third-party ms-swift framework rather than releasing a pretraining pipeline)

license

Apache-2.0(OSI)

Qwen3 Reranker ships Apache-2.0 weights and a repository that documents fine-tuning through a third-party framework, but no training corpus or data-construction scripts are released. That keeps it at open weights rather than a fully open release.

• https://huggingface.co/Qwen/Qwen3-Reranker-4B/raw/main/README.md recorded 2026-09-11

  Front matter reads license: apache-2.0 with base_model Qwen/Qwen3-4B-Base; the card links the QwenLM/Qwen3-Embedding repository and describes the 0.6B/4B/8B reranking line without naming a training corpus.

• https://huggingface.co/Qwen/Qwen3-VL-Reranker-2B/raw/main/README.md recorded 2026-09-11

  Front matter reads license: apache-2.0 for Qwen3-VL-Reranker-2B, base_model Qwen/Qwen3-VL-2B-Instruct.

• https://raw.githubusercontent.com/QwenLM/Qwen3-Embedding/main/README.md recorded 2026-09-11

  Repository README carries a Training section pointing at docs/training for the code and instructions to train Qwen3-Embedding models, alongside an Evaluation section.

Adoption is measured as Hugging Face downloads across the three text reranker checkpoints. The Qwen3-VL-Reranker line is scored as a separate product under the same text/vision-language split this map applies to Qwen3-Embedding, and community seq-cls conversions and GGUF/AWQ re-publishers are excluded.

• https://huggingface.co/api/models/Qwen/Qwen3-Reranker-4B recorded 2026-09-11

  downloads: 2546793; gated: false; private: false for Qwen/Qwen3-Reranker-4B.

• https://huggingface.co/api/models/Qwen/Qwen3-VL-Reranker-2B recorded 2026-09-11

  downloads: 1676336; gated: false; private: false for Qwen/Qwen3-VL-Reranker-2B.

Qwen3 Reranker is current-generation, multilingual and long-context, with strong published retrieval numbers across both text and visual-document tasks. It sits a step below Harrier-OSS, which holds a measured first place on the public MMTEB v2 leaderboard on the benchmark this category treats as primary.

• https://huggingface.co/jinaai/jina-reranker-v3/raw/main/README.md recorded 2026-09-11

  Third-party comparison table on the jina-reranker-v3 card places Qwen3-Reranker-4B at 61.16 BEIR, 67.52 MIRACL, 67.52 MKQA and 73.91 CoIR.

• https://huggingface.co/Qwen/Qwen3-VL-Reranker-2B/raw/main/README.md recorded 2026-09-11

  Model performance table reports Qwen3-VL-Reranker-8B at 79.2 MMEB-v2 retrieval average, 74.9 MMTEB retrieval, 83.6 JinaVDR and 66.7 ViDoRe v3, ahead of jina-reranker-m0 on VisDoc and ViDoRe.