WeMM-Embedding 2B: Choose 256 Dimensions Over 2048
Most people grab an embedding model and dump the full-size vector into their pipeline. Stop doing that.
Tencent’s WeMM-Embedding handles text, images, video, and documents. Its 2 billion-parameter version is ready for production.
The star feature is Matryoshka embeddings—a single pass can output 64, 128, 256, 512, 1024, or 2048-dimensional vectors.
Why smaller vectors matter
- Performance: At 256 dimensions, image and video tasks retain 98.7 % of the accuracy of the 2048-dim version.
- Speed: Smaller vectors let you rebuild indexes in a fraction of the time.
- Cost: You spend far less on storage and memory.
- Efficiency: Using 2048 dimensions when 256 suffices just burns resources.
The 2B model tops many larger competitors on the MMEB-v2 leaderboard. It’s already a production component—WeChat relies on it for search, e-commerce, and social media.
Documents need more space
Screenshots and other visual documents pack dense text and intricate layouts. Cutting dimensions hurts them more than photos.
How to test it
- Don’t overhaul your whole system at once.
- Pick 20 queries where your current setup stumbles—mix screenshots, product images, and document pages.
- Run those queries through the 2B model.
- If results improve, roll out the change. If not, you saved time and money.
Keep a scorecard per content type. Product photos may be fine at 256 dimensions, while PDF pages might need a higher minimum.
Source: https://dev.to/bean_bean/wemm-embedding-2b-chon-vector-256-chieu-vi-2048-2igd