Qwen3-Reranker

Open weights · Qwen, Alibaba · Updated 12 Oct 2026

The Qwen3-Reranker family (0.6B / 4B / 8B) are Apache 2.0 rerankers built on Qwen3, with a 32K context, 100+ languages and support for a task instruction. In Qwen’s own tests the 4B and 8B lead bge-reranker-v2-m3 and gte-multilingual-reranker on every suite they ran — English, Chinese, multilingual, long-document and code. Those are Qwen’s numbers on Qwen’s setup: validate on your own labels before replacing bge or a hosted API.

Model variants

All three: 32K context, 100+ languages, Apache 2.0.

For images, screenshots and video there is a separate family, Qwen3-VL-Reranker (2B / 8B, 32K context, 30+ languages, Apache 2.0). In Qwen’s own runs on ViDoRe v3, a visual-document retrieval benchmark, the 8B scores 66.7* and the 2B 60.8*; it is in the models table alongside LightOn-rerank, which also scores page images.

Recommendation: start with 4B on a single modern GPU. Drop to 0.6B only if memory is tight; move to 8B only when labelled eval shows a clear NDCG gain worth the cost.

Scores, and whose they are

Qwen’s own runs, from the model cards: the retrieval subsets of MTEB (English v2, Chinese v1, multilingual MMTEB and Code), MLDR for long documents and FollowIR for instruction-following, each reranking the top 100 candidates from Qwen3-Embedding-0.6B. MTEB-R is an English NDCG@10 average, but it isn’t the BEIR suite, so these numbers don’t share a column with other vendors’ BEIR figures.

One outside reference: Jina’s 2026 run puts Qwen3-Reranker-4B at 62.28* on 13 BEIR datasets — just under jina-reranker-v3.5 there, and ahead of it on multilingual, legal / medical and structured retrieval (see the Jina page). Until October 2026 this page compared Qwen’s scores with “bge (~60)” and “mxbai (~62)”; neither figure had a source, and the models table now shows what BAAI and mixedbread actually publish.

Quick start

sentence-transformers

pip install sentence-transformers

from sentence_transformers import CrossEncoder

model = CrossEncoder("Qwen/Qwen3-Reranker-4B")

# Default instruction: "Given a web search query, retrieve relevant
# passages that answer the query" — override it per task:
model = CrossEncoder(
    "Qwen/Qwen3-Reranker-4B",
    prompts={"support": "Prefer official help articles with step-by-step instructions"},
    default_prompt_name="support",
)

ranked = model.rank(query, documents)   # [{"corpus_id": i, "score": s}, ...]

The model cards load the rerankers straight into sentence-transformers’ CrossEncoder. The card also shows a plain transformers path (transformers ≥ 4.51) that scores yes/no logits through Qwen’s chat template — get that template wrong and quality drops without an error.

In a RAG pipeline

# Pseudocode — same two-stage pattern as any cross-encoder
candidates = vector_db.search(query, top_k=80)   # or hybrid
scores = qwen3_rerank(query, candidates)         # official template
top = sorted(zip(scores, candidates), reverse=True)[:8]
answer = llm.complete(context=top, query=query)

When to pick Qwen3

• Yes: you have GPU, need multilingual quality, and want open weights without per-call API cost.
• Yes: long passages matter — all three take 32K tokens.
• Yes: relevance needs steering. The rerankers take a task instruction; Qwen recommend writing one per task (in English, even for multilingual data) and report typical gains of 1–5%. See instruction-following rerank.
• Maybe not: CPU-only or edge — prefer bge-v2-m3, mxbai, or Jina tiny for demos.

Pros and cons

Other models