0xSero / GLM-5.3-Flash-EXL3-Spark

A mixed-precision EXL3 build of GLM-5.3-Flash sized to serve a 256k-token context with vision on a single NVIDIA DGX Spark (GB10, arm64, 128 GB unified memory): 12 of the 45 routed-expert MoE layers are quantized at 3.05 bpw for accuracy, the other 33 stay at 2.05 bpw for size…