Tencent Releases WeMM-Embedding

The Tencent WeChat Vision team has released WeMM-Embedding, a suite of general-purpose multimodal embedding models that integrates text, images, videos, and visual documents into a single vector space. The model weights and code have been open-sourced under the Apache 2.0 license.

Performance and Features

WeMM-Embedding is based on the Qwen3.5 architecture and is available in three sizes: 2B, 4B, and 9B. The smallest 2B model scored 77.9 on the MMEB-v2 benchmark, surpassing the previous top 8B open-source baseline, while the 9B model achieved 80.6, ranking first on the official leaderboard. Performance was validated in real-world service environments through 26 internal WeChat task benchmarks and 14 online A/B tests.

Training Strategy

Training consists of a first stage of large-scale alignment to ensure broad coverage, and a second stage of refinement training to enhance fine-grained matching capabilities using curated data.

• Data Construction: Six types of data, including weakly supervised pairs, captions, search, classification, multimodal QA, and graded relevance pairs, are converted into a unified pair format for training.
• Refinement Process: Semantic ID-based resampling is used to uniformize semantic distributions, MLLMs are utilized to refine data quality, and hard negatives are constructed to strengthen discriminative power.
• Representation Extraction: An <embedding> token is appended to the end of the input sequence, and the last-layer hidden state of this token is used as the output representation.