GLM-5.3-Flash was pruned from 288 to 144 routed experts per layer and quantized using an imatrix built from agentic coding sessions, producing a coder‑specialized model. The Q3‑quantized version (71 GB) matches the full 341 GB model’s capabilities, while the IQ2_XXS variant fits in 24 GB VRAM plus 32 GB RAM. It delivers comparable results on 26 Python SWE tasks from repositories such as Ansible and OpenLibrary.