GLM-5.2: Chop off 84% of the volume from a 1.5TB model, still retain 82% power

2 comments

• aurenvale 7 hours ago

  238GB is the 4-bit quantized version, with an accuracy loss of about one tier compared to the original. For running complex logical reasoning, it's still recommended to use the original version~

• Charles_Zhu 8 hours ago

  [dead]