跳转到主要内容

WeChat Vision Team Open-Sources Multimodal Model with Over One Billion Daily Calls

The WeChat Vision Team recently made a big move - they officially open-sourced the general-purpose multimodal embedding model WeMM-Embedding. This is not a lab toy; it has been running in WeChat's search and recommendation systems for a while, with over one billion calls per day. By open-sourcing this, they are sharing their core technology, with a clear goal: to allow more AI developers to stand on the shoulders of giants and create more interesting applications.

Three Versions, One for You

WeMM-Embedding is released in three different sizes: 2B, 4B, and 9B. Why so many versions? Simply put, to cater to developers with different needs. If you are building a lightweight application, the 2B version is sufficient - it runs fast and does not consume many resources. If you seek the ultimate performance, the 9B version provides stronger understanding capabilities. This flexible arrangement allows developers to choose based on their actual situation, without paying for unused performance.

Technically Robust, Yet Clear Approach

This model is indeed technically impressive. It adopts a unified architecture design and incorporates core techniques such as two-stage training and knowledge distillation. It sounds complex, but simply put: first, the model builds a solid foundation, then through a "master-apprentice" approach, the large model teaches the small model, enhancing understanding capabilities without making the model too heavy, ensuring efficient deployment.

Proven in Practice, Not Just on Paper

The most reassuring aspect is that WeMM-Embedding is not a model that performs well only on test sets but fails in real-world scenarios. It has been deployed in WeChat's recommendation and search systems, processing massive real-world requests daily, with over one billion calls per day. This practical experience has thoroughly validated the model's stability and reliability.

The Significance of Open Source: More Than Just Code

This open-source initiative is a strategic move for WeChat to build a multimodal AI technology ecosystem. For the industry, its significance is even greater. Previously, many teams wanted to use multimodal technology but either lacked the resources to train large models or could not find ready-made models that worked well. Now, by open-sourcing WeMM-Embedding, WeChat is handing the industry a key, enabling more teams to quickly get started and apply AI technology to more vertical fields such as healthcare, education, and e-commerce.

Promising Future

Multimodal AI is a major future trend, and this open-source is just the beginning. It is foreseeable that as more high-quality models like WeMM-Embedding become open-source, the pace of innovation in AI applications will accelerate. Ordinary people might soon use amazing applications driven by these technologies.

Key Points

  • Open-Source Model: WeMM-Embedding supports text and image inputs, available in 2B, 4B, and 9B versions.
  • Proven in Practice: Deployed in WeChat's search and recommendation systems, with over one billion daily calls.
  • Technical Highlights: Unified architecture, two-stage training, knowledge distillation, balancing effectiveness and efficiency.
  • Industry Impact: Open source will empower AI developers and promote the application of multimodal AI in more fields.