Tencent released WeMM-Embedding, a family of universal multimodal embedding models that align text, images, videos, visual documents, and interleaved inputs in a shared embedding space (Huggingface). The family includes 2B, 4B, and 9B parameter variants trained through large-scale multimodal alignment followed by refinement using curated data and fine-grained relevance supervision (Huggingface).
On public benchmarks, the 2B variant already surpasses the previous leading 8B open-source baseline, while the 9B variant achieves a state-of-the-art score of 80.6 on MMEB-v2 (Huggingface). In production, WeMM-Embedding has been deployed at scale across WeChat's recommendation and search applications, including WeChat Channels, Official Accounts, Moments, and e-commerce services, with consistent improvements across 14 online A/B tests (Huggingface). For commerce teams, these open-source models provide a foundation for efficient cross-modal search and personalization without the cost of training embeddings from scratch.
The models and code are available on GitHub, enabling broader adoption for retrieval, recommendation, classification, and agentic systems in commerce applications (Huggingface).