Jina Embeddings v5 Omni Nano

Jina Embeddings v5 Omni Nano is a lightweight multilingual multimodal embedding model with 1.004B parameters built on the jina-embeddings-v5-text-nano text backbone, extended with a fine-tuned SigLIP2 Base vision encoder (from Qwen3.5-0.8B) for image and video and a Whisper-large-v3 audio encoder (from Qwen2.5-Omni-7B), connected via trained cross-modal projectors. It supports 8K context length and produces 768-dimensional embeddings with Matryoshka representation learning (down to 32 dimensions). It uses last-token pooling and supports retrieval, text-matching, clustering, and classification tasks across text, image, video, audio, and PDF inputs.