Tencent Releases WeMM-Embedding for Multimodal Retrieval

Tencent's WeChat Vision team released WeMM-Embedding, a family of multimodal embedding models in 2B, 4B, and 9B parameter sizes, supporting text, images, videos, visual documents, and interleaved inputs. The models support Matryoshka dimensions for adjustable vector widths, with reported average scores of 77.9, 79.2, and 80.6 on MMEB-v2 for the respective sizes. The repository includes inference examples, serving instructions, and evaluation code.

Tencent’s WeChat Vision team has published WeMM-Embedding https://github.com/Tencent/WeMM-Embedding , a family comprising 2B, 4B and 9B embedding models. Each variant supports text, images, videos, visual documents and interleaved multimodal inputs; audio is not supported. For developers, the immediate consequence is a single repository containing model options, inference examples, serving instructions and evaluation code for retrieval across the supported inputs. The models obtain embeddings from the last-layer hidden state at a dedicated <embedding token, then apply L2 normalization. The project presents this as a unified representation method across its supported input types rather than separate embedding formats for each modality repository https://github.com/Tencent/WeMM-Embedding . All three variants support reduced Matryoshka dimensions. The 2B model offers widths from 64 to 2,048, the 4B model from 64 to 2,560 and the 9B model from 64 to 4,096. For a supported smaller width, the repository instructs users to truncate the full vector and normalize it again repository https://github.com/Tencent/WeMM-Embedding . Inference examples cover Transformers and Sentence Transformers. Tencent recommends transformers==5.2.0 for inference and reproducibility because newer releases may differ in preprocessing, while the serving instructions list vLLM 0.27.0 and SGLang 0.5.9 as tested versions repository https://github.com/Tencent/WeMM-Embedding . On MMEB-v2, which covers 78 datasets, Tencent reports average scores of 77.9, 79.2 and 80.6 for the 2B, 4B and 9B models respectively. The table uses Hit@1 for image and video tasks and NDCG@5 for visual-document tasks repository https://github.com/Tencent/WeMM-Embedding . The broader MMEB-v3 evaluation contains 190 tasks, including the 78 MMEB-v2 tasks alongside text, agent, audio and MCMR tasks. Tencent reports V3-All scores of 56.0 for the 2B model, 58.2 for 4B and 59.5 for 9B. All three receive zero for audio because unsupported tasks are assigned zero, and the repository includes the MMEB-v3 evaluation code used for the reported results repository https://github.com/Tencent/WeMM-Embedding . Analysis: Adjustable vector width is a more directly testable design choice than the headline benchmark ranking. Teams can compare supported dimensions within one model family, but the repository substantiates performance retention with one specific result: on MMEB-v2, the 2B model at 256 dimensions retains 98.7% of its full-dimensional image and video performance. That does not resolve retention for text, visual documents, interleaved inputs or a production corpus. The leaderboard figures should likewise be treated as project-reported evidence until independently reproduced repository https://github.com/Tencent/WeMM-Embedding .