Tencent Hunyuan open source Hy4: 770 billion parameter multi-modal model, all Hy series are open source

📅 2026-11-20

Abstract:

On November 20, Tencent Hunyuan announced the official open source Hy4 model, and simultaneously released the Hy4 System Card, disclosing the model's security assessment method, test results and known risks. Developers can currently obtain relevant models and deploy them through GitHub, Hugging Face, ModelScope and GitCode.


Public information shows that Hy4 is a multi-modal model using a hybrid expert architecture that supports text, image and video understanding. The total number of parameters of its language model body is 770 billion, each Token activates approximately 49 billion parameters, and supports a context window of up to 1 million Tokens. In addition to the standard instruction model, this open version also includes the FP8 quantization model and the pre-trained basic model.

The System Card released together with the model weights is mainly used to explain the safety evaluation of Hy4, and provides a reference for developers to understand the boundaries of model capabilities and potential risks. Tencent Hunyuan said that this move continues its practice of simultaneously disclosing security assessment information during the model release process.

With the opening of Hy4, the Hy series of large language models currently released by Tencent Hunyuan have been open source, involving Hy3 preview, Hy3, Hy4 preview, Hy4 and Hy4 flash versions. Among them, Hy4 flash has previously opened up model weights and is positioned in application scenarios that require high inference efficiency and deployment costs.

Related tags

Related articles

Comments

0/500
Captcha (click to refresh)
No comments yet