Ovis-Embedding puts text, images, video and audio in one retrieval space

The Alibaba team reports state-of-the-art results across five multimodal embedding suites using a shared backbone rather than separate modality towers.