Ask or search anything...

Jina AI's `jina-embeddings-v4` is a 3.8 billion-parameter model designed for universal multimodal and multilingual retrieval, featuring a unified VLM backbone that minimizes the modality gap between text and images. The model achieves state-of-the-art results on the Jina-VDR and ViDoRe benchmarks for visually rich document retrieval and shows competitive performance across diverse text and multimodal tasks.

Jina AI's jina-reranker-v3 model implements a "last but not late interaction" (LBNL) mechanism, enabling a Qwen3-0.6B transformer to concurrently process queries and multiple documents for listwise reranking. This method achieved a 61.85 nDCG@10 on the BEIR benchmark, exceeding the performance of significantly larger rerankers, and showed strong capabilities across diverse languages and specialized domains.

The research from Jina AI and Weaviate introduces "Late Chunking," a method that generates contextual chunk embeddings by first encoding an entire document with a long-context embedding model and then applying chunking and pooling. This approach consistently improves retrieval accuracy (nDCG@10 by 2.70-3.63% on BeIR benchmarks) for neural search and RAG systems while avoiding the computational costs of LLM-based contextualization.

Jina AI introduces jina-embeddings-v3, a 570-million-parameter multilingual text embedding model that leverages Task LoRA for efficient adaptation to diverse tasks. It achieves superior performance on MTEB English and multilingual benchmarks, especially in long-context retrieval up to 8192 tokens, surpassing larger LLM-based embeddings.

Jina AI researchers introduce ReaderLM-v2, a compact 1.5B parameter language model that outperforms GPT-4 by 15-20% in HTML-to-Markdown/JSON conversion tasks while handling documents up to 512K tokens, demonstrating that specialized smaller models can excel at structured content extraction through innovative three-stage training.

Jina AI introduced Jina Embeddings v2, the first open-source general-purpose text embedding model designed to process documents up to 8192 tokens, a 16x increase over typical limits. The model achieves competitive performance on standard benchmarks and demonstrates superior results on long-document specific tasks, outperforming proprietary models like OpenAI's `ada-embeddings-002` on the LoCo benchmark.