Tencent releases WeMM-Embedding-4B universal multimodal embedding model

Tencent has released WeMM-Embedding-4B, a universal multimodal embedding model built on Qwen3.5 that accepts text, images, videos, visual documents, and interleaved inputs to return 2,560-dimensional L2-normalized embeddings.

The model supports independent encoding of any content subset and is available via pip install with transformers and sentence-transformers libraries. It also offers integration with vLLM 0.27.0 for pooling inference and SGLang 0.5.9 for precise embedding interpolation.

WeMM-Embedding-4B demonstrates strong performance across 78 datasets from the MMEB-v2 benchmark, utilizing Hit@1 for image/video tasks and NDCG@5 for visual documents.