Tencent's WeChat Vision Team has open-sourced WeMM-Embedding, a family of general-purpose multimodal embedding models, with both code and weights posted simultaneously to GitHub and Hugging Face. The technical report is filed as arXiv 2608.24053, and the code is released under the Apache 2.0 license.
An embedding model's job is to compress content into a vector, so that similar things sit close together in vector space — retrieval, deduplication, and recommendation all lean on it. WeMM-Embedding covers more ground than most peers: text, images, video, visual documents, and arbitrary interleaved combinations of these inputs all run through the same representation. Audio is not supported for now.
Three Sizes, Top of the Leaderboard
The series ships in three sizes: 2B, 4B, and 9B. Across the 78 datasets in MMEB-v2, the 2B model scores 77.9 overall (79.6 on images, 70.8 on video, 80.7 on visual documents), the 4B model reaches 79.2, and the 9B model hits 80.6 — first place on the official leaderboard, ahead of every open and closed model listed.
The head-to-head comparisons put the position in sharper focus. The 2B version beats Qwen3-VL-Embedding-2B by 4.7 points and DME-2B by 3.1 points, while also slightly edging out Qwen3-VL-Embedding-8B, a model with four times its parameter count. In this kind of task, the weight of training recipe and data quality has already overtaken raw parameter scale.
Training runs in two stages: large-scale multimodal alignment first, then refinement on curated data. The refinement stage adds fine-grained relevance supervision and cross-scale knowledge transfer — distilling what the larger models learn down into the smaller ones, which is part of how the 2B model manages to punch above its weight class.
Trimming the Dimensions
All three sizes support Matryoshka representation learning, with output dimensions selectable anywhere from 64 up to 2048 or 4096. The official figures show that trimming to 256 dimensions on MMEB-v2 still retains 98.7% of the full-dimension performance on image and video tasks.
This matters more to engineering teams than to the leaderboard. Storage and retrieval overhead in a vector database scales roughly with dimension count, so cutting from 2048 down to 256 shrinks the index to an eighth of its size, with distance calculations at recall time dropping accordingly. Previously, achieving this meant training a separate small model or bolting on a dimensionality-reduction layer; now the same set of weights can simply be truncated to the first few segments. Under the hood, the embedding is drawn from the final layer's hidden state at the dedicated <embedding> token position, then L2-normalized.
Already Running in Production
The technical report mentions an internal evaluation set spanning 26 tasks, along with 14 groups of live A/B experiments across WeChat's businesses, all showing consistent gains. Those figures can't be independently reproduced, but they indicate the model hasn't stopped at leaderboard-chasing — the sheer scale of WeChat's Search, Channels, and Official Accounts content retrieval is itself a stress test.
The limitations are spelled out clearly too: no audio input support, and the evaluation code samples video at 64 frames. Handling long-form video or joint audio-visual retrieval would require an additional layer built on top.
For teams building multimodal RAG in China, the practicality of the 2B tier carries more weight than the 9B model topping the chart. Rough math suggests a 2B model paired with 256-dimensional output can index tens of millions of image-text pairs on a single GPU — a completely different cost structure from relying on closed-source APIs. The Apache 2.0 license on the code also keeps the bar for commercial use low.
Sources: Tencent/WeMM-Embedding open-source repository, arXiv technical report 2608.24053, Hugging Face model page, CocoLoop; MMEB-v2 scores by tier, the Matryoshka dimension range, and the count of live A/B experiment groups follow the official repository and technical report.