Tencent Hunyuan Open Source Hy4 Quantization Compression Model "Hy4 preview Lightweight Version"

Tencent Hunyuan has released a lightweight version of Hy4 preview, which compresses the model weights from 1.5 TB to approximately 214 GB using its self-developed Sherry sparse ternary quantization algorithm and MIX-STQ1_0 mixed-precision layer-by-layer quantization strategy. The quantized version of Hy4 preview performs close to the original BF16 in tasks such as long text understanding, multi-turn context retrieval, and encoding assistance.