Wechat unveils open-source multimodal embedding model wemm-embedding

On September 4, 2026, WeChat AI announced the open-sourcing of its general-purpose multimodal embedding model, WeMM-Embedding, which is already widely deployed in WeChat's recommendation and search systems. The model, which processes billions of requests daily, comes in three versions: 2B, 4B, and 9B, supporting text, images, videos, visual documents, and mixed multimodal inputs. WeMM-Embedding is built on the Qwen3.5 multimodal architecture, allowing images, videos, and text to be processed together from the start. The 9B version achieved a score of 80.6 on the MMEB-v2 benchmark, ranking first, while the 2B version scored 77.9, surpassing previous leading models. The model is used across various platforms including video, live streaming, and e-commerce, with daily usage reaching the 10 billion level. WeChat AI reported improvements in multiple tasks, including classification, search, and cross-domain matching, with the 2B version showing a significant increase in performance.