Qwen3-VL-Reranker-8B

Name: Qwen3-VL-Reranker-8B
Author: Qwen

by Qwen

8B cross-encoder reranker for text, image, screenshot, and video retrieval

52Kdl/month

165likes

8.8Bparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

Qwen/Qwen3-VL-Reranker-8B

Feature URI

mixpeek://reranker@v1/qwen3_vl_reranker_8b_v1

Overview

Qwen3-VL-Reranker-8B is the larger sibling of the 2B variant, built on the Qwen3-VL-8B-Instruct vision-language backbone. As a cross-encoder, it jointly attends to query and document tokens with full cross-attention, producing a single relevance score per pair. It handles text, images, screenshots, and video frames in both query and document positions, supporting 30+ languages and inputs up to 32K tokens.

The 8B variant consistently outperforms the 2B model across most multimodal reranking benchmarks, making it the best choice when precision matters more than latency. On Mixpeek, it serves as a second-stage reranker that rescores candidates retrieved by a fast bi-encoder embedding model, dramatically improving top-K precision for cross-modal search.

Architecture

Cross-encoder architecture built on Qwen3-VL-8B-Instruct. Jointly processes query and document tokens with full bidirectional cross-attention. Outputs a scalar relevance score per pair. Supports text, images, screenshots, and video frames as both query and document inputs. 32K token context window. 30+ language support.

Mixpeek SDK Integration

// Reranking is a retriever STAGE in Mixpeek, not an ingest-time extractor.
// The rerank stage runs a cross-encoder inference service; the shipped default
// is BAAI/bge-reranker-v2-m3. Pointing it at Qwen3-VL-Reranker-8B means registering that
// model as a custom reranker plugin and naming it in feature_uri, which is an
// Enterprise path. Stage contract read from GET /v1/discovery/stages.
const retriever = await mx.retrievers.create({
  namespace_id: "my-namespace",
  retriever_name: "search-then-rerank",
  stages: [
    {
      stage_name: "candidates",
      stage_id: "feature_search",
      parameters: { limit: 100 },
    },
    {
      stage_name: "rerank_results",
      stage_id: "rerank",
      parameters: {
        inference_name: "BAAI__bge_reranker_v2_m3",
        query: "{{INPUT.query}}",
        document_field: "content",
        top_k: 10,
      },
    },
  ],
});

Capabilities

Cross-modal reranking across text, image, screenshot, and video
Full cross-attention for higher precision than bi-encoder models
32K token context for long-document reranking
30+ language support including CJK
Consistently outperforms 2B variant on most benchmarks

Use Cases on Mixpeek

Two-stage multimodal retrieval: fast embedding search followed by high-precision reranking

Visual document search: rerank candidate PDF pages by cross-attending to query and page screenshot

Video scene reranking: rescore candidate scenes by cross-modal relevance to text queries

Benchmarks

Dataset	Metric	Score	Source
Multimodal reranking (MMEB-V2)	nDCG@10	Best in class (8B)	Qwen, 2026: arxiv,2601.04720
Visual document reranking	nDCG@10	Outperforms 2B variant	Qwen, 2026: arxiv,2601.04720