Overview
WeMM-Embedding-2B is an open multimodal embedding model from Tencent’s WeChat Vision team, built on Qwen3.5. It maps text, images, video, visual documents and interleaved multimodal content into a shared representation space. (huggingface.co)
Capabilities and uses
- Produces a 2,048-dimensional, L2-normalized embedding. Matryoshka dimensions of 64, 128, 256, 512, 1,024 and 2,048 are supported by truncation and renormalization. (github.com)
- Intended for multimodal understanding, cross-modal retrieval, search and recommendation. Tencent reports use of the model family in WeChat search and recommendation applications. (tencent.github.io)
- Tencent reports an average score of 77.9 for the 2B model on MMEB-v2, across 78 datasets. (github.com)
Limitations and deployment
- Audio input is not supported. The model produces embeddings rather than conversational text. (huggingface.co)
- Smaller embeddings trade representation size for potential performance loss; the reported 256-dimensional image and video results retain 98.7% of full-dimensional performance on MMEB-v2, not a guarantee for every task. (github.com)
- Tencent provides model weights and deployment examples. The model is licensed under Apache 2.0, while third-party components retain their respective license terms. (huggingface.co)
Publication date
The technical report’s first version was submitted on August 25, 2026. This is the verified report submission date, not necessarily the model’s separate release date. (arxiv.org)
Pricing
As of November 16, 2026, no official model-specific input or output API price was verified on Tencent’s mainland China or international sites. CNY/RMB and USD pricing, billing units and any separate output charge are therefore unverified. Self-hosting may incur infrastructure costs.