Dynamic Beating AI News Flash: Tencent WeChat Vision Team has open-sourced WeMM-Embedding, releasing three multimodal embedding models in one go: 2B, 4B, and 9B. Based on Qwen3.5, these models can transform text, images, videos, and visual documents into vectors for search, recommendation, and content matching. All models are open source under the Apache 2.0 license.
Interestingly, Qianwen (Qwen) has not yet launched Qwen3.5-Embedding. Currently, the official versions are Qwen3-Embedding and Qwen3-VL-Embedding. Tencent, on the other hand, has taken the lead in developing the new generation Embedding based on Qwen3.5.
The performance is quite impressive. The WeMM-Embedding 2B model achieved a score of 77.9 on MMEB-v2, slightly higher than the score of 77.8 achieved by the Qwen3-VL-Embedding 8B model. The 9B version further improved to 80.6. Both models support cross-modal retrieval, such as text-to-video and image-to-text. However, with the WeMM model integrated with the Qwen3.5 base, higher performance was achieved with a smaller model.
This model set has already been put into practical use within WeChat. Tencent has stated that it has been applied in WeChat Official Accounts, Video Accounts, e-commerce recommendations, and WeChat Search, which also covers Moments search.
