Qwen3-Reranker

Alibaba's Apache 2.0 instruction-aware rerankers (0.6B, 4B, 8B): 100+ languages, 32K context, built on Qwen3.

In short. Qwen3-Reranker re-sorts search results. It reads a question and a candidate passage together and judges whether the passage answers it, so the best passages move to the top. It comes in 0.6B, 4B and 8B sizes under Apache 2.0, and it covers more than 100 languages and code.

Summary

Qwen3-Reranker is the reranking half of Alibaba's Qwen3 Embedding series, which the Qwen team released on 5 June 2025. It is built on the dense Qwen3 models with a cross-encoder design and LoRA fine-tuning, so it reads a query and a document together as one text. The card's Transformers example turns the answer into a score by comparing the model's probabilities for the words "yes" and "no", and Sentence Transformers can load it as a CrossEncoder. It supports more than 100 languages, including programming languages, and a 32K-token context, although the card's example caps inputs at 8,192 tokens. The reranker accepts a custom instruction, and the card says instructions typically improve results by 1% to 5% and recommends writing them in English. Qwen trained the rerankers directly on high-quality labeled data and skipped the weakly supervised stage used for the embedding models.

In Qwen's table, which reranks the top 100 candidates retrieved by Qwen3-Embedding-0.6B, the 0.6B reranker scores 65.80 on MTEB-R, 71.31 on CMTEB-R, 66.36 on MMTEB-R, 67.28 on MLDR and 73.42 on MTEB-Code. The 4B scores 69.76, 75.94, 72.74, 69.97 and 81.20, and the 8B scores 69.02, 77.45, 72.94, 70.19 and 81.22. On the same table BGE-reranker-v2-m3 scores 57.03, 72.16, 58.36, 59.51 and 41.38. The 4B beats the 8B on MTEB-R and on the FollowIR instruction-following test, 14.84 against 8.05, so the larger model is not automatically the better one. These are Qwen's own runs.

The license is Apache 2.0, and the BF16 weights are 1.19 GB, 8.04 GB and 16.38 GB. Loading needs Transformers 4.51 or later, and ggml-org publishes a Q8_0 GGUF of the 0.6B model for llama.cpp. In the 30 days to 11 November 2026, Hugging Face counted about 2.63 million downloads of the 4B model, 1.29 million of the 0.6B and 79,000 of the 8B.

Good for

  • Reranking the top 50 to 100 results of any retrieval system, in 100+ languages and in code
  • Tuning the ranking with a custom instruction for your domain or language, which Qwen says gains 1% to 5%
  • The 0.6B model for cheap serving, the 4B for the best balance in Qwen's table and the 8B for the highest Chinese and multilingual scores
  • Commercial products, thanks to the Apache 2.0 license

Watch out for

  • Every score is Qwen's own, measured on candidates from Qwen3-Embedding-0.6B
  • The 4B beats the 8B on some tests, so try both on your data
  • A cross-encoder must read every query and candidate pair, so delay grows with the candidate list
  • It reads text only, so use Qwen3-VL-Reranker for images and video

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Qwen3-Reranker. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Reranker-0.6B BF16 596M 1.19 GB 1.37 GB Laptop or phone
Reranker-4B BF16 4.02B 8.04 GB 9.25 GB Laptop or phone
Reranker-8B BF16 8.19B 16 GB 19 GB Gaming GPU

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Qwen3-Reranker?
Qwen3-Reranker re-sorts search results. It reads a question and a candidate passage together and judges whether the passage answers it, so the best passages move to the top. It comes in 0.6B, 4B and 8B sizes under Apache 2.0, and it covers more than 100 languages and code.
Can I use Qwen3-Reranker commercially?
Yes. Qwen3-Reranker is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Qwen3-Reranker open source?
Qwen3-Reranker is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Qwen3-Reranker on my own computer?
Yes. The smallest variant, Reranker-0.6B, needs about 1.37 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Qwen3-Reranker released?
Qwen3-Reranker was released by Alibaba Qwen on June 5, 2025.
How long a context window does Qwen3-Reranker have?
Qwen3-Reranker supports a context window of up to 32K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
BGE Reranker v2 m3 BAAI 568M 8K Apache 2.0 (Permissive) 2.61 GB Laptop or phone
Jina Reranker v3.5 Jina AI 597M 128K CC BY-NC (Non-commercial) 0.46 GB Laptop or phone
Qwen3-Embedding Alibaba Qwen 600M – 7.57B 32K Apache 2.0 (Permissive) 1.38 GB Laptop or phone
Qwen3-VL-Reranker Alibaba Qwen 2.13B – 8.77B 32K Apache 2.0 (Permissive) 4.9 GB Laptop or phone