Tencent open-sources Hy4 preview 770B-parameter AI model

Tencent has released Hy4 preview, an open-source 770-billion-parameter Mixture-of-Experts (MoE) AI model. The system features 49 billion active parameters and a one-million-token context window, designed to support complex tasks such as long-horizon software engineering, scientific research, game development, and document-heavy office workflows.

In performance benchmarks, Hy4 preview reportedly outperformed models such as GPT-5.6 Sol on the SWE Atlas Refactoring and SWE-bench Pro tests. The model was trained using specialized data from software engineers, finance analysts, and security experts to enhance its ability to plan, debug, and verify work across extended sessions.

Tencent has also provided quantized versions of the model to reduce hardware requirements. Using a selective quantization method (STQ1_0), the model can be compressed significantly, with versions requiring approximately 214GB of VRAM, allowing for more accessible deployment while maintaining high performance levels.