malaiwah/GLM-5.3-Flash-TR3-8bpw
Image-Text-to-Text • 166B • Updated • 115 • 2
GLM-5.3-Flash — measured quants & fidelity
TR3 trellis quants of GLM-5.3-Flash with receipt-backed KL measurements on a sealed panel, plus the fidelity registry.
Note K8 - 0.012384 nats, 1.66x better than official FP8 at the same 331 GB footprint
malaiwah/GLM-5.3-Flash-TR3-6bpw
Image-Text-to-Text • 127B • Updated • 198 • 4Note K6 - 0.013723 nats sealed over 5 bitwise-identical runs, 254 GB
malaiwah/quant-fidelity-registry
Viewer • Updated • 157 • 226 • 1Note Comparable fidelity measurements across quants and model families
malaiwah/GLM-5.3-Flash-fidelity-suite-v1
Updated • 169Note The quality reference: BF16/FP8 hidden states, 10.48M positions, all receipts
malaiwah/GLM-5.3-Flash-calibration-activations-v1
Viewer • Updated • 1 • 60Note 147 GB MoE activations + router logits for calibration-aware quantization
malaiwah/GLM-5.3-Flash-TR3-partsbin-v1
Updated • 41Note The parts bin: K6+K8 payload stores under one transform seed — mixed-precision builds become offline CPU assembly