Tencent releases lightweight version of hy4 preview model with enhanced efficiency

On September 1, 2026, Tencent's Hybrid Intelligence team announced the release of a lightweight version of the Hy4 preview model, reducing its model weights from nearly 1.5TB to approximately 214GB. The Hy4 preview model, which was released and open-sourced on August 28, 2026, is a new generation of MoE (Mixed Expert) large language model with a total parameter count of 770B and an activation parameter count of 49B. The compression was achieved through two core technologies: Sherry sparse ternary quantization algorithm and MIX-STQ1_0 hybrid precision strategy. These technologies enabled the model to maintain stable performance across mainstream tasks, with results comparable to the original BF16 model. The lightweight version is now available on Hugging Face and GitHub, supporting multiple inference frameworks.