Tencent’s WeChat Vision team released WeMM-Embedding, a family of three open embedding models (2B, 4B, and 9B parameters) that place text, images, video, page scans and mixed inputs of several kinds at once inside one shared vector space. The weights, code, and a technical report are posted on GitHub and Hugging Face under an Apache 2.0 license.
The pitch is consolidation. Instead of running separate encoders for text search, image retrieval, and document lookup, a team can query one model and compare results directly in the same embedding space. That matters most for retrieval-augmented generation systems that mix formats: a support bot searching product manuals, screenshots, and chat transcripts together needs exactly this kind of shared representation, and right now most RAG stacks stitch it together with separate encoders and a lot of glue code.
WeMM-Embedding also supports Matryoshka representation learning, which lets a developer truncate a full-size embedding down to a smaller dimension (as low as 64) and renormalize it without retraining. According to the GitHub repository, the 2B model holds 98.7 percent of the image and video scores it posts at full width when squeezed to 256 dimensions on MMEB-v2, a meaningful storage and latency win if the number holds up in independent testing.
On benchmarks the team reports in its own technical report, the 9B model scores 80.6 average on MMEB-v2’s 78 datasets, ahead of Alibaba’s Qwen3-VL-Embedding 8B at 77.8 and well past GME 8B at 59.2. On the newer MMEB-v3 suite of 190 tasks, WeMM-Embedding 9B scores 59.5 against Qwen3-VL-Embedding 8B’s 53.5. Every one of these numbers comes from Tencent’s own paper, not an independent leaderboard, and the repository does not link to third-party reproductions.
The MMEB-v3 numbers also expose the model’s edge case. That suite includes 11 audio retrieval tasks, and WeMM-Embedding scores exactly 0.0 on all of them because, as the repository says without hedging, the models do not take audio at all yet. The model’s headline average absorbs that zero rather than excluding it, which understates how well it does on the modalities it actually covers, image, video, and visual documents, but also means any team searching audio content will need a separate pipeline regardless of the strong text and image scores.
WeMM-Embedding is a data point in a broader pattern: unified multimodal retrieval is becoming the plumbing layer under multimodal RAG, and a major Chinese consumer platform shipping that plumbing as open weights, rather than as a paid API, extends the open-weights race beyond chat models into the infrastructure that sits underneath them. Teams building document or media search on top of CLIP-era embeddings should benchmark WeMM-Embedding’s 2B checkpoint against their current stack before the next procurement cycle, since Matryoshka truncation could cut both storage costs and latency without a retraining pass.
Released by WeChat’s Vision team at Tencent on GitHub, repository and technical report published 2026.
