Tencent Compresses Hy4-Preview Model to 200GB GGUF with 98% Performance Retained
Tencent has successfully compressed its next-generation Hy4-preview model from a massive 1.5TB down to approximately 200GB in the GGUF format. Despite this drastic reduction in size, the compressed model retains about 98% of its original performance. This compression significantly lowers the hardware barrier, making it feasible to run a frontier-class large language model on consumer or prosumer-grade hardware rather than requiring enterprise-level GPU clusters. It accelerates the accessibility of local AI deployment for advanced models. The Hy4-preview is a mixture-of-experts (MoE) model featuring 770 billion total parameters, 49 billion active parameters, and a context window exceeding 1 million tokens. The conversion to GGUF format allows for efficient inference and fast loading on local systems.
## BACKGROUND
Large Language Models (LLMs) typically require massive amounts of VRAM to run because their weights are stored in high-precision formats like FP16. Quantization is a technique that converts these weights to lower-precision formats, drastically reducing the model's memory footprint. GGUF is a popular binary file format designed by the llama.cpp community to optimize the loading and execution of these quantized models on consumer hardware.