Tencent Releases WeMM-Embedding, a Multimodal Embedding Model Family
Tencent has released WeMM-Embedding, a family of open-weights multimodal embedding models with 9B, 4B, and 2B parameters built on Qwen3.5. These models can process text, images, videos, visual documents, and interleaved multimodal inputs to generate unified embeddings. This release provides the open-source community with powerful tools for advanced Retrieval-Augmented Generation (RAG) systems that need to search across diverse media types simultaneously. By supporting interleaved inputs, it enables more complex search queries that combine text and visual elements seamlessly. The WeMM-Embedding-9B model outputs a 4,096-dimensional L2-normalized embedding, which simplifies similarity calculations to dot products in vector databases. However, the model family currently does not support audio inputs.
## BACKGROUND
Embedding models convert unstructured data like text or images into dense numerical vectors that capture semantic meaning. Multimodal embeddings project different data types into a shared vector space, allowing a text query to retrieve relevant images or videos. L2-normalization scales these vectors to a unit length of one, meaning that a simple dot product calculation between two vectors yields their cosine similarity, making search operations highly efficient.