WeMM-Embedding is a family of universal multimodal embedding models developed by the WeChat Vision Team at Tencent, providing unified representations for text, images, videos, visual documents, and interleaved multimodal inputs. It ships in three sizes -- 2B, 4B, and 9B parameters -- each trained with Matryoshka Representation Learning so a single model produces embeddings at multiple dimensions (from 64 up to 2048 for the 2B, 2560 for the 4B, and 4096 for the 9B) that can be truncated and re-normalized without retraining; on MMEB-v2, the 2B model at 256 dimensions retains 98.7% of its full-dimensional image and video performance. Embeddings are read from the last-layer hidden state at a dedicated <embedding> token position followed by L2 normalization. The repo reports evaluation results on MMEB-v2 (78 datasets, Hit@1 for image/video, NDCG@5 for visual documents) and MMEB-v3 (190 tasks spanning image, video, visual-document, text, agent, and MCMR categories), where all three model sizes lead their size class on the aggregate columns against the baselines listed in the README's own tables, including VLM2Vec, VLM2Vec-V2, GME, and Qwen3-VL-Embedding. Audio input is explicitly not supported in the current release. Inference is supported through Hugging Face Transformers and Sentence Transformers, and serving is documented for both vLLM (0.27.0) and SGLang (0.5.9), with one-command wrapper scripts for each.