Ovis-Embedding is an omni-modal embedding model family built on a shared multimodal backbone that natively encodes text, image, video, and audio into one representation space rather than assembling separate modality towers. Techniques include homogeneous-source batch sampling for informative negatives, focal loss on hard examples, similarity-based embedding distillation from complementary experts, and inference-time low-rank feature decomposition for compact, flexible-dimension embeddings. It reports state-of-the-art results across the MMEB-v3, MMEB-v2, MVEB, MAEB, and RTEB benchmarks.
