WeChat Open-Sources WeMM-Embedding Multimodal Vector Model Suite
WeChat AI has open-sourced WeMM-Embedding, a universal multimodal embedding model suite available in 2B, 4B, and 9B parameter sizes. Built on a unified Qwen3.5 multimodal backbone, the 9B variant achieved a top score of 80.6 on the MMEB-v2 benchmark, while the 2B model scored 77.9, surpassing previous 8B open-source baselines. Unlike many academic models, WeMM-Embedding is battle-tested at industrial scale, serving over 1 billion daily production calls across WeChat's search and recommendation infrastructure. Open-sourcing high-performing models proven in production gives developers access to state-of-the-art multimodal retrieval capabilities for video, document, and text processing. The suite handles text, images, video, visual documents, and arbitrary interleaved inputs within a single unified vector representation space. Across 26 internal evaluation benchmarks at WeChat, the 2B model improved average performance from 60.9 to 72.0 across search, classification, cross-domain matching, and content relevance tasks.
## BACKGROUND
Multimodal embedding models map different types of media—such as images, video clips, and text—into dense vectors within a shared vector space, enabling cross-modal search and retrieval. Benchmarks like MMEB-v2 evaluate these models across diverse retrieval, grounding, and reasoning tasks to measure how effectively they connect semantic concepts across mediums.