Tencent opens the weights of the Hy4 and WeMM-Embedding model families
Tencent released the weights of Hy4 preview and put the model into its own apps, while the WeChat Vision team opened the code and weights of the WeMM-Embedding family for matching text, images and video. The releases cover separate jobs in chat and multimodal retrieval. Published performance results are Tencent’s own measurements, and the Hy4 release provides model weights despite the company’s open-source label.
Artificial Intelligence··Evening
Hy4 runs selected experts from a large model
Tencent released Hy4 preview with 770 billion total parameters and 49 billion active parameters per token. Its context window exceeds 1 million tokens, and the weights are available on Hugging Face under Apache 2.0 terms. The model is also accessible in Yuanbao, ima, WorkBuddy and CodeBuddy, as well as through Tencent Cloud TokenHub and OpenRouter. According to the company’s model card, most of the 78-layer backbone contains 256 routed experts and 1 shared expert, with the top 8 routed experts activated for each token.[1]
The WeChat Vision team’s WeMM-Embedding family represents and matches text, images, video and other content types. Tencent released 2, 4 and 9 billion parameter sizes, with code, evaluation tools and weights available through a public repository and model pages. The team says the models already run in WeChat Channels, Official Accounts, Moments and its e-commerce services. In Tencent’s own results, the 9 billion parameter version ranks first among listed models on MMEB-v2 and MMEB-v3. That position comes from the company’s evaluation and has no independent measurement.[2]
The releases expose different parts of the stack
Hy4 and WeMM-Embedding show two release paths within the same company’s model distribution. Tencent describes Hy4 as open source; the released package consists of weights, while the model also runs in Tencent consumer apps and cloud services. WeMM-Embedding exposes code, evaluation tools and weights together, giving developers access to a broader part of the family for multimodal search, retrieval and recommendation work. Published performance claims for both models come from Tencent’s measurements. The public WeMM-Embedding repository also carries the model code and evaluation tools.[1], [2]
