WeChat Open-Sources WeMM-Embedding Multimodal Model Tested at Industrial Scale
WeChat AI has open-sourced WeMM-Embedding, a family of universal multimodal embedding models featuring 2B, 4B, and 9B parameter variants. The model handles text, image, video, visual document, and interleaved inputs, while achieving the top rank on the MMEB-v2 benchmark. WeMM-Embedding is proven at massive industrial scale inside WeChat, handling over 1 billion daily API calls across Moments search, Channels recommendations, Official Accounts, and e-commerce. Releasing a top-performing embedding model that is battle-tested in heavy enterprise workloads offers open-source developers a robust foundation for multimodal retrieval and search systems. Built on the Qwen model family, WeMM-Embedding incorporates Matryoshka Representation Learning (MRL), enabling a single trained model to generate nested embeddings across flexible vector dimensions for optimized storage and retrieval. The models are available on Hugging Face and GitHub, supporting seamless integration using the SentenceTransformers library.
## BACKGROUND
Multimodal embedding models map diverse media types—such as text, images, and videos—into a unified vector space where algorithms can perform semantic similarity search across formats. MMEB-v2 (Massive Multimodal Embedding Benchmark) is a key benchmark designed to evaluate multimodal embedding models across tasks including retrieval, classification, moment grounding, and question answering.