jina-reranker-m0

jina-reranker-m0 is a 2.4B-parameter multimodal multilingual reranker based on a vision-language model architecture. It is the first reranker to handle visual documents (text, figures, tables, charts) across multiple languages, achieving state-of-the-art performance on ViDoRe (91.02 NDCG-5) and multilingual long-document retrieval (MLDR). The model supports text-to-text, text-to-image, image-to-text, and mixed-modality reranking.

The model is built on a decoder-only VLM architecture (2.4B parameters) combining a DFN CLIP ViT vision encoder with a Qwen2 language decoder. Unlike traditional cross-encoders that process text only, the VLM backbone natively handles both text and image inputs, enabling reranking of documents that contain visual content (scanned PDFs, infographics, slide decks, tables with embedded figures). The 10K token context window accommodates up to 768 tokens per image (processed as 768×28×28 patches). Training uses a multi-modal contrastive objective on diverse document types across multiple languages. The decoder-only architecture enables listwise reranking and opens possibilities for document deduplication and ranking score explainability via attention mechanisms — capabilities unavailable with encoder-only architectures.

In text-to-text reranking, the model scores 58.95 NDCG-10 on BEIR, outperforming jina-embeddings-v3 (55.81) and bge-reranker-v2-m3 (56.51). For multilingual content, it achieves 66.75 NDCG-10 on MIRACL (18 languages). On the MLDR benchmark for long documents, it scores 59.83 NDCG-10 across 13 languages. Code retrieval reaches 63.55 NDCG-10 on CoIR. The model shines in visual document retrieval: 91.02 NDCG-5 on ViDoRe, and 43.92 average on Winoground (visio-linguistic compositional reasoning). Some modality combinations (e.g., image-to-image) are supported in zero-shot mode without specific training data.

The model is accessible via Jina API, AWS, Azure, GCP marketplaces, or locally through Hugging Face. When using the API, pass text strings, base64 images, or image URLs — new users receive 10M free tokens. The model supports up to 10K input tokens with up to 768 tokens per image. For optimal results, the model performs best on text-to-text, text-to-image, image-to-text, and text-to-mixed-modality tasks; image-to-image is zero-shot. The decoder-only architecture enables capabilities beyond simple reranking: mixed-modality reranking, listwise reranking, document deduplication, and attention-based ranking explainability. Use this model when your documents contain visual content (scanned PDFs, charts, infographics) that text-only rerankers cannot handle. For pure text reranking at lower cost, use jina-reranker-v3.5.