Training-free MoE expert pruning by functional replaceability. Four backbones at 25% and 50% expert removal, no recovery training.
-
Nickyang/GLM-4.7-Flash-Razor-17B-A3B-E32of64
Text Generation • 17B • Updated • 207 • 2 -
Nickyang/GLM-4.7-Flash-Razor-24B-A3B-E48of64
Text Generation • 24B • Updated • 186 • 2 -
Nickyang/Hy3-Razor-154B-A21B-E96of192
Text Generation • 154B • Updated • 116 • 1 -
Nickyang/Hy3-Razor-226B-A21B-E144of192
Text Generation • 226B • Updated • 109 • 1
