WeMM-Embedding is an open multimodal embedding model released by Tencent's WeChat Vision Team, built for engineers who need one vector representation across text, images, video, and documents for search, retrieval, and recommendation. Open-sourced on September 4, 2026, it is the same model that powers recommendation and search across WeChat at more than one billion daily calls.
Core Features
- Three sizes, 2B, 4B, and 9B, all based on the Qwen3.5 multimodal architecture.
- Unified encoding of text, images, video, visual documents, and interleaved inputs into one shared dense vector space.
- Matryoshka representation so teams can truncate to smaller dimensions with minimal quality loss; 2B reaches 2048 dimensions, 9B hits 4096.
- Top benchmark results: 9B scores 80.6 on MMEB-v2 for the number-one spot, while 2B at 77.9 beats prior 8B open models.
- Covers retrieval, classification, question answering, visual grounding, and agent memory or GUI retrieval.
Use Cases
- Multimodal RAG systems that match queries against images and documents together.
- Recommendation and ranking features inside content or commerce apps.
- Cross-modal search across video, image, and text catalogs.
- Agent tool and memory retrieval where the agent must find the right context fast.
Pricing
WeMM-Embedding is free and open-weight, with an effective license cost of $0. Tencent released the model weights, inference code, and evaluation tools on Hugging Face and GitHub, so there is no license fee; you pay only for the compute you run. It already serves WeChat Channels, Live, Official Accounts, e-commerce, and Moments at production scale.
Our Take
Best for teams building multimodal search or retrieval who want a proven, production-grade embedding model they can self-host. The trade-off is that 9B needs real GPU memory, so start with 2B and scale only when accuracy demands it. See related AI Search tools.
