WeChat AI officially announced on September 4 that it will open-source the general multimodal embedding model WeMM-Embedding, which includes three versions: 2B, 4B, and 9B. It supports text, images, videos, visual documents, and any interleaved multimodal input.
According to a WeChat employee @ Ke Cun Xiao Jiang, the multimodal model WeMM-Embedding open-sourced by the WeChat Visual Team essentially allows the system to understand text, images, and videos simultaneously, placing them in the same semantic space for comparison and retrieval.
Already applied in scenarios such as Moments search and video recommendation
He revealed that this model has been widely used within WeChat—Moments search, video recommendation, public account recommendation, and WeChat e-commerce all use it, with a daily call volume of 1 billion times. On the international authoritative benchmark MMEB-v2, the model ranked first, surpassing all previously submitted open-source and closed-source models.
Currently, related resources of WeMM-Embedding have been made available: the paper was published on arXiv, the code repository is hosted under the Tencent/WeMM-Embedding project on GitHub, and the model set is also available on Hugging Face.
