WeChat open-sources WeMM-Embedding: 2B outperforms 8B Qwen3-VL, 9B tops MMEB-v2

WeChat has open-sourced WeMM-Embedding, a multimodal embedding model based on Qwen3.5 in 2B/4B/9B sizes. The 2B version achieves 77.9 on MMEB-v2, surpassing Qwen3-VL-Embedding 8B (77.8), while the 9B scores 80.6 to lead the benchmark. The model accepts interleaved text, image, video, and visual document inputs, and can output two vectors (pure video and video+captions) in one forward pass. Matryoshka dimensions from 64 to 4096 allow 2B to retain 98.7% performance at 256 dim. Training uses hundreds of millions of (instruction, source, target, hard negative, graded relevance) quintuples, curated subset with density resampling and MLLM correction, and a two-stage process with InfoNCE, CoSENT, reranker supervision, and embedding distillation from a frozen 9B teacher. Deployed on WeChat Channels, Official Accounts, Moments, and e-commerce recommendation/search, 14 A/B tests all showed positive results.