✦
Summary
Ovis-VL-Embedding-9B maps text and visual inputs into a shared representation space for multimodal retrieval. Built from Qwen3.5-9B, it removes the language-modeling head and extracts a final-layer retrieval embedding without separate modality-specific projection heads.
About Ovis-VL-Embedding-9B
Search and RAG developers can index mixed visual documents, images, and video with one model rather than maintaining isolated search spaces. Core capabilities include Text embeddings, Image embeddings, Visual-document retrieval, and Video embeddings.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Interleaved multimodal input, Shared representation space, Nine-billion-parameter backbone, and Apache-2.0 weights.