WeChat Open-Sources Multimodal Embedding Model WeMM-Embedding

WeChat announced today that it has open-sourced its general-purpose multimodal embedding model, WeMM-Embedding, available in 2B, 4B, and 9B versions. The model supports text, images, videos, visual documents, and any interleaved multimodal inputs. Built on the Qwen3.5 multimodal architecture, it processes different data types in a unified framework. On the MMEB-v2 benchmark, the 9B version ranks first with 80.6 points, while the 2B version scores 77.9, surpassing the previous leading 8B open-source model.

WeMM-Embedding is now widely deployed in WeChat's recommendation and search systems, covering Channels, live streaming, official accounts, e-commerce, and Moments, with over one billion daily calls. The embeddings are used for candidate retrieval, ranking feature construction, user sequence modeling, and cross-domain content understanding. In 26 internal task benchmarks, the 2B version improved from 60.9 to 72.0 compared to similar-sized baselines, with gains across classification, search, cross-domain matching, and other tasks.