Qwen3-VL Reranker (multimodal)
Scores relevance across modalities, text against images and back
$1 in free credits
Overview
Qwen3-VL-Reranker-8B is the reranking half of the Qwen3-VL retrieval suite, built on Qwen3-VL-8B-Instruct. It takes a query and document pair where either side may be text, an image, a screenshot, video or a mixture, and returns a relevance score. Qwen intends it as the second stage after Qwen3-VL-Embedding recall. It covers 30+ languages; the checkpoint supports 32K context, and this deployment scores up to 4,096 tokens per query and document pair.
Strengths
- Reranks across modalities, not only text
- Designed to pair with qwen3-vl-embedding
- 30+ languages
Use cases
- Refining multimodal knowledge-base search
- Text-to-image and image-to-text relevance
Good to know
Used through knowledge-base search; there is no public rerank endpoint. 8B parameters against bge-reranker's ~568M makes it far costlier per candidate. Worth it only when the documents really are multimodal.
API access
One OpenAI-compatible API for every model in the catalog, plus the CLI and the function SDK inside workflows. Sign up and try $1 of inference free, then get $10 more once you've spent $5, and these requests work:
# qwen3-vl-reranker is a reranker model; the model API has no rerank route. # A knowledge base reranks with the reranker paired with its embedding model; # a search can name another in reranker_model. curl https://model.graphn.ai/v1/models \ -H "Authorization: Bearer $GRAPHN_API_KEY"
graphn model get qwen3-vl-reranker
@function
async def search(kb_id: str, query: str) -> list:
return await kb.search(kb_id, query, top_k=5, rerank=True)