WeMM-Embedding

WeMM-Embedding is a family of universal multimodal embedding models from Tencent's WeChat Vision team, released 2026-08-25. It maps text, images, videos, visual documents, and interleaved multimodal inputs into a single shared, L2-normalized embedding space, so one model can index and retrieve across formats instead of running separate encoders. The family ships in three sizes, 2B, 4B, and 9B, all built on Alibaba's open Qwen3.5 checkpoint, and is released under Apache 2.0.

Compare Qwen3-VL-Embedding if you wanted the prior Alibaba multimodal embedding row, Qwen3-Embedding for the text-only twin, and Voyage-3-Large if you preferred a hosted enterprise embedding API.

Key Features

• Truly multimodal: Each model accepts text, images, videos, and scanned/visual documents plus interleaved mixes, and returns a fixed-dimension vector via last-token pooling at a dedicated <embedding> token.
• Matryoshka dimensions: The 9B model exposes 64, 128, 256, 512, 1024, 2048, and 4096; the 4B exposes up to 2560; the 2B up to 2048. Truncate to a smaller dimension and keep most of the quality (the 2B retains ~98.7% of full-dimension image and video quality at 256 dims).
• Qwen3.5 base: Finetuned from Alibaba's Qwen3.5-9B / 4B / 2B, which lets WeChat Vision reuse an open foundation instead of training from scratch.
• Serving paths: First-class support for vLLM (0.27+) and SGLang (0.5.9) pooling backends, plus the wemm_sentence_transformers loader for a familiar encode API.
• Apache 2.0: Code, weights, and parameters are public; third-party components keep their own licenses.

Limitation: audio input is not supported, so the MMEB-v3 audio score is zero by design. The flagship 9B model needs a sizable GPU to be useful.

Use Cases

• Multimodal RAG: A single WeMM index over documents, images, and clips lets you retrieve across formats with one vector store instead of several.
• Visual document search: PDFs, slides, and scanned pages fall into one embedding space with text.
• Video retrieval: 64-frame sampling in the MMEB-v3 code path indexes clips alongside images and text.
• Teams already on Qwen3.5: Because the base is Qwen, adapting from a Qwen3.5 workflow is straightforward.

Pricing

WeMM-Embedding is open weights under Apache 2.0, so there is no per-token price. The cost is your own compute.

No Inference Provider hosts it yet on 2026-08-25, so plan to self-host via vLLM or SGLang.

Getting Started

1. pip install torch transformers==5.2.0 "qwen-vl-utils[decord]==0.0.14" sentence-transformers==5.7.0 "accelerate>=1.1.0".
2. Load with load_wemm_sentence_transformer("tencent/WeMM-Embedding-9B", device="cuda:0").
3. Encode text, { "image": ... }, and { "video": ... } items together, setting truncate_dim to your target Matryoshka dimension.
4. Scale with vllm serve ... --runner pooling or the SGLang patch in the repo.

First-party start: Hugging Face collection and technical report.

Frequently Asked Questions

Does it support audio?

No. Text, image, video, and visual documents are supported; audio is assigned a zero in MMEB-v3.

Is it open source?

Yes. Tencent-authored code and weights are Apache 2.0, on Hugging Face and GitHub.

What's the embedding dimension if I don't truncate?

The 9B returns 4096, the 4B returns 2560, and the 2B returns 2048. You can truncate to any listed Matryoshka dimension.

Alternatives

• Qwen3-VL-Embedding: Prior Alibaba multimodal embedding snapshot.
• Qwen3-Embedding: Text-only Qwen embedding twin.
• Voyage-3-Large: Hosted enterprise embedding and rerank API.

Tips

1. Start with the 2B and truncate to 256 dims if you are GPU-limited; it keeps ~98.7% of full-dimension image and video quality.
2. Use the dedicated <embedding> token and last-token pooling, not mean pooling, to match the training setup.
3. Recheck the leaderboard tables in the technical report before quoting MMEB numbers; the audio sub-score is zero by design.

Conclusion

WeMM-Embedding is an open, Apache-2.0 multimodal embedding family from Tencent that unifies text, image, video, and visual documents in one space, built on a Qwen3.5 base. Start at the Hugging Face collection, then decide whether the Qwen3.5 lineage or a hosted option like Voyage-3-Large fits your retrieval stack.