WeChat open-source multimodal embedding model WeMM-Embedding

SummaryAI generated

WeChat has open-sourced the multi-modal embedding model WeMM-Embedding. This model supports embedding representations of various modal data such as text, images, audio, and video. It can efficiently convert unstructured multi-modal data into dense representations in vector space. It is widely used in natural language processing, computer vision, speech recognition, and other scenarios, providing important foundational components for multi-modal large language models and intelligent systems.

Related eventsRELATED EVENTS

Key entitiesKEY ENTITIES

MMEB-v2Qwen3.5WeChatWeMM-Embedding

Coverage · reports per dayLANGUAGE SPLIT

Entity relations

SignalsSIGNALS

Keyword heat

WeChat1
WeMM-Embedding1
Qwen3.51
MMEB-v21

All reports (1)SOURCES

↗

D DoNews·快讯 zh 2026-09-04 08:00

On September 4, 2026, WeChat AI announced the open-source of the general multi-modal embedding model WeMM-Embedding. This model comes in three versions: 2B, 4B, and 9B, and is based on the Qwen3.5 architecture. The 9B version ranked first on the MMEB-v2 list with a score of 80.6, while the 2B version achieved a score of 77.9, surpassing the previous leading 8B open-source model. Currently, the model is deployed in core scenarios such as WeChat Video Accounts, live broadcasts, and official accounts, with an average daily call volume of over 1 billion times. It is used for retrieval, sorting, user modeling, and cross-domain understanding. Related papers, code, and models have been published on arXiv, GitHub, and Hugging Face.