WeChat Multi-modal Embedding Technical Report
Listen to the summary
Uses a voice available on your device
Audio options
On this page 4 sections
Related concepts 6 concepts
Key Takeaways
- The 9B model reaches a top score of 80.6 on the MMEB-v2 benchmark.
- The 2B variant outperforms the previous 8B open-source baseline on MMEB-v2.
- The models show consistent performance improvements in 14 online A/B tests at WeChat.
- The model architecture is built on the Qwen3.5 backbone to handle multimodal inputs.
Summary & Methodology Analysis
WeMM-Embedding introduces a suite of three universal multimodal embedding models with 2B, 4B, and 9B parameters. These models utilize the Qwen3.5 backbone, which is natively designed for multimodal input, to process various data types into a unified vector representation. By mapping these inputs into a dense vector space, the models enable downstream retrieval tasks across diverse datasets. The training pipeline involves a two-stage approach: a broad initial alignment phase followed by a refinement stage that incorporates hard-negative mining and embedding distillation to enhance vector quality.
Interactive System Flowchart
Cross-Examination & FAQs
A deeper dive clarifying mechanics, constraints, and baseline evaluations.
Q1. What is the primary contribution of WeMM-Embedding?
It provides a set of multimodal embedding models that achieve high performance on standardized benchmarks and production applications.
Q2. Are these models suitable for production use?
Yes, they have shown consistent gains across 14 online A/B tests within WeChat applications.
Q3. Do these models support audio data?
No, the models currently do not support audio input, and audio-related tasks are scored as zero.
Q4. How do the models perform relative to existing open-source baselines?
The 2B model outperforms the leading 8B open-source baseline on MMEB-v2, and the 9B model ranks first on the official leaderboard.
Q5. What benchmarks were used to evaluate these models?
The models were evaluated on MMEB-v2 and MMEB-v3, the latter of which includes complex text retrieval, agent-centric tasks, and audio evaluation.
Q6. Does using a reranker improve performance for these models?
Empirically, the authors found that reranking candidates does not consistently improve performance across all multimodal tasks.
Q7. What is the parameter range of the model series?
The series includes three models with 2B, 4B, and 9B parameters respectively.
Q8. What foundation model backbone does WeMM-Embedding use?
The models are built on the natively multimodal Qwen3.5 backbone.
Q9. Are there any specific limitations mentioned regarding task performance?
The paper notes that reranker supervision does not consistently yield improvements and that the models lack support for audio input.