Tencent Launches WeMM-Embedding — A Multimodal Model for Query Processing
Tencent Launches WeMM-Embedding — A Multimodal Model for Query Processing
Tencent has unveiled the WeMM-Embedding, a new multimodal model that processes around 1 billion queries daily. The model supports text, images, and video, offering flexibility in dimensionality for various tasks.
Tencent has launched WeMM-Embedding — a multimodal model that processes approximately 1 billion queries per day within Weixin.
The model works with text, images, and videos, taking into account the order in which they appear.
The 9B version ranked first in MMEB-v2 with a score of 80.6.
The smaller 2B version retains 98.7% quality even at a dimensionality of 256.
Moreover, the dimensionality can be reduced to 64 for fast searching or increased to 2048 for more accurate ranking — without retraining the model.
Now, WeMM-Embedding is available in open access.
Why it matters
AnalysisThis model opens new possibilities for processing multimedia queries, which can significantly enhance user experience across various applications. The open access to the model will also foster the development of AI technologies.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Tencent releases Hy4: an open-weight model challenging DeepSeek
Tencent has released Hy4, a multimodal model that scored 75.2 on APEX-Agents in internal tests, ahead of DeepSeek V4.1 Flash-1017. The company plans to publish the model weights.
Google Gemini Pro (Model with Infinite Context)
The flagship working model from Google DeepMind featuring a record context window of over 2 million tokens. Capable of analyzing entire books, lengthy videos, and massive codebases in a single query.
Native Audio: Direct Speech-to-Speech Processing
The new generation of native multimodal models (GPT-4o Advanced Voice, Gemini Live) processes sound waves directly without the intermediate step of converting audio to text (STT) and back (TTS). This allows the model to perceive sarcasm, fear, laughter, whispers, and interrupt conversations on the fly with minimal latency.
MIT researchers show GPT-6 Astra can improve other AI models by editing their weights
In an experiment, Astra raised Qwen3.5-0.8B's accuracy on MATH-500 from 7.9% to 48.1% by changing only 0.016% of the model's weights.