Tencent released its next-generation large language model, Hy4. Versus the Hy4 preview, the model expands pretraining and post-training (fine-tuning) scale, adds vision capabilities, and strengthens instruction following, long-text comprehension, long-range agent performance and real-world productivity tasks. Hy4 contains 770B parameters and uses a mixture-of-experts architecture. It also introduces ComputerUse, enabling the model to operate both command-line and graphical interfaces for complex

Tencent released its next-generation large language model, Hy4. Versus the Hy4 preview, the model expands pretraining and post-training (fine-tuning) scale, adds vision capabilities, and strengthens instruction following, long-text comprehension, long-range agent performance and real-world productivity tasks. Hy4 contains 770B parameters and uses a mixture-of-experts architecture. It also introduces ComputerUse, enabling the model to operate both command-line and graphical interfaces for complex tasks. In tests across four practical scenarios Hy4 ranks among the leading global open-source models, and Tencent says model weights will be open-sourced soon.