WeMM-Embedding-9B
WeMM-Embedding-9B is built on Qwen3.5-9B. It maps text, images, visual documents, videos and mixed inputs into 4,096-dimensional normalized vectors.
Model design and access
The supplied code can encode each modality alone or combine items in one input. Query and document methods are provided through Sentence Transformers. The output is an embedding for retrieval or similarity, rather than a generated answer.WeMM-Embedding-9B official model documentation
Practical use
Practical guidance. Prioritize difficult cross-modal queries in evaluation. Keep frame sampling consistent when searching video, because missed events cannot be recovered by the embedding model.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| License | Apache 2.0; third-party components retain their licensesWeMM-Embedding-9B official model documentation |
| Audio | Not supportedWeMM-Embedding-9B official model documentation |
| Vector output | 4,096 dimensions; L2-normalizedWeMM-Embedding-9B official model documentation |
Related models
- WeMM-Embedding-2B
Tencent · WeMM-Embedding
A 2B Qwen3.5-derived embedding model that maps text, images, documents and video to 2,048-dimensional vectors.
Launch date not confirmedDownloadable - WeMM-Embedding-4B
Tencent · WeMM-Embedding
A 4B Qwen3.5-derived embedding model that maps text, images, documents and video to 2,560-dimensional vectors.
Launch date not confirmedDownloadable