WeMM-Embedding is available in three versions: 2B, 4B, and 9B. The model supports a variety of multimodal inputs, including text, images, videos, and visual documents, as well as any interleaved own inputs.
The model's primary function is to enable systems to understand text, images, and videos simultaneously by placing them within a single semantic space for comparison and retrieval.
The model is already integrated into several WeChat services. It is used in Moments search, Channels recommendation, Official Accounts recommendation, and WeChat e-commerce.
The system handles a massive scale of operations, with 1 billion calls per day. It has also achieved the top position on the MMEB-v2 international benchmark, surpassing all previously submitted open-source and closed-source models.
