Key Info

Tencent Hy4 Preview's weights are compressed from ~1.5TB to 214 GiB for a 770B-parameter model, aiming to preserve capabilities and practical inference speed.

Highlights

  • Reduces model size by ~86% from 1.5TB to 214 GiB.
  • Maintains useful capabilities and inference speed.
  • Discussed in a Zhihu article by Tencent Hunyuan member yghstill.