Tencent’s HunYuan team said it compressed Hy4 preview model weights from 1.5TB
to about 214GB. The team reports the compressed model retains near-original
long-document comprehension, matches the full model on multi-turn long-context
retrieval, and shows only a slight decline in math ability; overall performance
is described as within acceptable ranges. The model covers coding assistance,
tool calling, long-document processing and routine Q&A.