WeMM-Embedding — Multimodal Embedding Model

WeMM-Embedding is a general-purpose multimodal embedding model unifying text, image and video representations to support cross-business retrieval inside the WeChat ecosystem.

Approach

• Backbone: Qwen3-VL.
• Deep Fusion module integrating multi-level semantic representations.
• Deduplicated InfoNCE loss to remove false negatives.
• Hierarchical sampler supporting arbitrary modality mixing during training.

Results

• MMEB-V2 0.7523 — open-source #1.
• UVRB 0.686 for video retrieval — open-source #1.
• +2.6% improvement on Chinese image–text retrieval.