inference-optimization/GLM-5.3-Flash-MEP50

GLM-5.3-Flash - 50% Expert Pruned

50% of the MoE experts pruned by router-weight magnitude using compressed-tensors.

• —Base model: zai-org/GLM-5.3-Flash
• —Sparsity: 50% of routed experts removed per layer
• —Layers pruned: all 43 MoE layers (body layers 3-44)
• —MTP layer: retained exactly as-is
• —Shared experts: untouched
• —Vision tower: untouched (dense ViT MLP, no MoE experts)

Reproduction