GitHub radar
WeMM-Embedding: A Generalized Solution for Text, Image, and Video Search
Multimodal embedding models developed by Tencent’s WeChat Vision for joint text, image, video, and visual document search.
The WeChat Vision team at Tencent has introduced a series of universal multimodal embedding models: WeMM-Embedding. These models unify text, image, video, visual document and interleaved multimodal information into a single vector space for cross-modal search. For instance, they allow users to retrieve relevant video snippets or documents based on text queries, or find images similar to user-uploaded images. There are three variants of WeMM-Embedding, each with 2B, 4B and 9B parameters. All the model weights are released on Hugging Face platform, named as tencent/ . You can easily use these models with standard transformers or sentence-transformers. It also supports production deployment with vLLM and SGLang. With a Matryoshka embedding mechanism, it enables you to project a larger embedding vector (64–4096) to a smaller one without re-training, by sacrificing some performance for saving storage and inference time.
Why a vibe-coder should care
When you build a search system to index a multimedia library, a question-answer system to answer questions based on documents, or a multimodal recommender system, you have to maintain multiple embedding models and write glue codes to integrate them for different modalities. With WeMM-Embedding, you only need a single model and a single embedding space to cover all your needs, which greatly reduces the system complexity and allows you to serve multi-modal requests with a single API.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up WeMM-Embedding-2B from https://github.com/Tencent/WeMM-Embedding — follow the README, download the model from Hugging Face (tencent/WeMM-Embedding-2B), and show me an example: get embeddings for a text query and an image, then compare them with cosine similarity.
The 2B version needs a Mac with 16+ GB of RAM. The 9B version needs an NVIDIA GPU with 20+ GB of VRAM — you can't run it on a regular laptop.
Open on GitHub▌ More finds