Decode · tok/s
30
Prefill
—
tok/100W
12.5
glm-5.3 flash on one GB10, 262K context, running at 30 tok/s. quant: 0xSero's M288 mosaic 12L. EXL3 2.05 bpw base + 12 layers of 3-bit routed experts.
First-person one GB10. Distinct from the stock EXL3 2.05bpw 32.32 pair: this checkpoint mixes 3-bit routed experts. Rank is the quoted 30.
