Tencent Releases WeMM-Embedding-9B, a Multimodal Embedding Model Built on Qwen3.5

Tencent has released WeMM-Embedding-9B, a universal multimodal embedding model built on the Qwen3.5 base architecture. The model accepts text, images, video, visual documents, and interleaved multimodal inputs, and outputs a 4,096-dimensional L2-normalized embedding vector. Audio input is not supported.

Benchmark performance

According to Tencent's technical report, WeMM-Embedding-9B scores 80.6 average on MMEB-v2 across 78 datasets (image, video, and visual-document tasks), which the company states surpasses comparable models including Qwen3-VL-Embedding 8B (77.8) and DME-Medium 9B (78.4). Breakdown scores: 81.9 on image tasks (Hit@1), 74.3 on video tasks (Hit@1), and 83.3 on visual-document tasks (NDCG@5).

On the broader MMEB-v3 benchmark, which spans 190 tasks including 53 text tasks, 47 agent tasks, 11 audio tasks, and multi-context multi-round (MCMR) tasks, WeMM-Embedding-9B posts a 59.5 average — the highest reported among models compared in the release. It scores 48.8 on text (NDCG@5), 51.0 on agent tasks, and 49.3 on MCMR (Hit@1). The model does not support audio tasks and receives a score of zero on that subset, consistent with its stated capability limits.

Tencent also released smaller 2B and 4B variants of WeMM-Embedding, which score 77.9 and 79.2 respectively on MMEB-v2, and 56.0 and 58.2 on MMEB-v3 — all following the same architecture family.

Technical details

The model supports Matryoshka embeddings, allowing users to truncate the 4,096-dimensional output to smaller dimensions (e.g., 256) with renormalization, useful for reducing storage and compute costs in retrieval systems. It is compatible with the Hugging Face Transformers library (version 5.2.0), Sentence Transformers (5.7.0+), vLLM (0.27.0), and SGLang (0.5.9) for serving.

WeMM-Embedding-9B is a finetune of Qwen/Qwen3.5-9B-Base and is one of 693 finetunes derived from that base model listed on Hugging Face. The model has been downloaded 2,375 times in the past month and is published under the Apache License 2.0, permitting commercial use.

The accompanying technical report, "WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report" (arXiv:2608.24053), lists authors Junjie Zhou, Ke Mei, Lei Li, Tianyi Wang, Fengyun Rao, and Jing Lyu, and was published seven days prior to this release.

What this means

Multimodal embedding models are becoming a competitive front distinct from generative LLMs, as retrieval systems increasingly need to index images, video, and documents alongside text in a shared vector space. Tencent's release, tied to its WeChat platform team, targets this infrastructure layer directly rather than chat or agent use cases. The Apache 2.0 license and support across major serving frameworks (vLLM, SGLang) make it straightforward for developers to self-host, though pricing for any hosted API access is not yet disclosed since no inference provider currently deploys the model. The reported benchmark gains over Qwen3-VL-Embedding and other open models are Tencent's own figures from its technical report and have not been independently verified.