Wianmu-Embedding

Wianmu is a model for universal multimodal embeddings. It is WeMM-Embedding with ears.

WeMM-Embedding embeds text, images and video, but it can't hear. Wianmu keeps WeMM and the audio encoder of Qwen2.5-Omni-3B frozen, and trains only a small connector that maps audio into WeMM's embedding space. Text, image and video embeddings are exactly WeMM's, and audio now lands right next to them. The name comes from WeMM + Tianmu, hence Wianmu.

Models

Wianmu-3B	Wianmu-5B	Wianmu-10B (soon)