jina-reranker-m0: Multilingual Multimodal Document Reranker

**What the feature is**
jina-reranker-m0 is a next-generation multilingual, multimodal reranking model designed to bridge the gap between text queries and visually rich documents. Built on the Qwen2-VL-2B architecture, it moves beyond traditional text-only cross-encoders to a decoder-only vision-language model (VLM). The model can process and rank a diverse range of inputs, including:
- Visually Rich Documents: PDFs, screenshots, and images containing text, tables, figures, and complex layouts.
- Multilingual Content: Support for over 29 languages across text and visual domains.
- Long Context: A massive 32K token context window, specifically optimized for long-form documents and complex code searching.
- Mixed Modality: The ability to rank a collection that contains both raw text and images simultaneously against a single query.

**Value proposition**
The jina-reranker-m0 provides a unified solution for modern retrieval pipelines by solving three critical pain points:

- Elimination of the "Modality Gap": Unlike previous CLIP-based models where images and text clustered separately, m0 creates a truly unified representation. This ensures that an image and a text document of equal relevance receive comparable scores, enabling seamless hybrid retrieval.
- State-of-the-Art (SOTA) Performance: It delivers market-leading results on visual benchmarks like ViDoRe (Visual Document Retrieval) and M-BEIR, while simultaneously outperforming previous models in specialized text tasks like CoIR (Code Information Retrieval).
- Operational Simplicity: By consolidating visual and textual reranking into a single 2.4B parameter model, developers can replace fragmented pipelines, which previously required separate OCR engines and text-only rerankers with a single API call that understands both pixels and prose.