Qwen3-Reranker

Open weights · Qwen / Tongyi · Updated 25 Jun 2026

The Qwen3-Reranker family (0.6B / 4B / 8B) is the headline open-weight reranker story of 2026: multilingual, long-context friendly, and frequently topping community open-weight leaderboards. Treat published MTEB-R / BEIR-style numbers as directionally strong — always validate on your own labels before replacing bge or a hosted API.

Model variants

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 & honesty

Public write-ups often cite MTEB-R / multilingual rerank suites in the ~70+ range for mid/large Qwen3 sizes, with 8B at the top of open-weight tables. Those figures are not guaranteed to match the classic BEIR 18-dataset averages we list for bge (~60) or mxbai (~62).

• Different suites, languages, and preprocessing → different numbers.
• On some English-only product benches, compact models (e.g. ModernBERT-GTE) can beat a 4B model on Hit@1 while being far cheaper.
• Our comparison table marks Qwen scores with * for this reason.

Quick start

Hugging Face / transformers

# Check the model card for the recommended stack (transformers / vLLM / SGLang).
# IDs (examples):
#   Qwen/Qwen3-Reranker-0.6B
#   Qwen/Qwen3-Reranker-4B
#   Qwen/Qwen3-Reranker-8B

from sentence_transformers import CrossEncoder  # if card supports CE API

# Prefer the loading recipe on the official model card — Qwen3 may use
# a chat-style scoring template rather than a plain CrossEncoder pair.
model_id = "Qwen/Qwen3-Reranker-4B"
# model = CrossEncoder(model_id, max_length=8192)  # only if supported
# scores = model.predict([(query, d) for d in docs])

For production serving, many teams use vLLM / OpenAI-compatible /v1/rerank endpoints. Follow the current Qwen docs for the exact prompt format — wrong templates silently destroy quality.

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 / long context matter and you can afford 4B–8B memory.
• Maybe not: CPU-only or edge — prefer bge-v2-m3, mxbai, or Jina tiny for demos.
• Maybe not: instruction-shaped policies — see instruction-following rerank.

Pros and cons

Other models