Overview
WeMM-Embedding-9B is a roughly 9-billion-parameter universal multimodal embedding model developed by Tencent’s WeChat Vision team. It represents different types of content in a shared vector space rather than generating conversational answers. (github.com)
Capabilities and uses
- Accepts text, images, video, visual documents and interleaved multimodal inputs.
- Supports Matryoshka embedding dimensions of 64, 128, 256, 512, 1,024, 2,048 and 4,096. Smaller vectors can reduce storage requirements, with a potential quality trade-off.
- Applicable to cross-modal retrieval, search, recommendation, classification and retrieval-augmented applications. The technical report describes deployment in WeChat search and recommendation scenarios. (github.com)
- Tencent reports an average score of 80.6 on MMEB-v2 across 78 datasets. This is a reported benchmark result, not a guarantee of performance on a particular application. (github.com)
Availability and limitations
Tencent publishes model weights and code under Apache License 2.0; third-party components retain their own licenses. The model can be self-hosted. Audio input is not currently supported. The repository recommends Transformers 5.2.0 for reproducibility because preprocessing may differ in newer versions. Deployment requires suitable computing resources. (huggingface.co)
Release date
The supplied release date is August 25, 2026. The first version of the official technical report was submitted on that date; this does not independently establish a separate model-weights release timestamp. (arxiv.org)
Pricing
As of November 16, 2026, an official hosted API price specifically for WeMM-Embedding-9B could not be verified on Tencent’s domestic or international pricing pages. Input and output charges, billing units and currencies are therefore unverified. The pages list prices for other embedding models, which should not be applied to this model. Self-hosting may incur infrastructure costs. (cloud.tencent.com)