GLM-5.2-NVFP4-REAP-469B

REAP-pruned GLM-5.2 NVFP4. Original 763.6B params → 469.2B params (61.4% retained).

Calibration Data (ALL sources, 16.1B tokens)

• GLM-5.2 original 22k paper-style calibration
• GLM-5.2 owned SFT + formatting (3,914 samples)
• GLM-5.2 remaining 16k (3,999 batches)
• GLM-5 22k layerwise observations (0xSero/glm5-layerwise-reap-observations)
• GLM-5.1 batchparallel observations (0xSero/glm-5-special)

Pruning

• 256 → 156 experts per layer (100 pruned)
• 7,500 experts removed of 19,200 total
• REAP saliency = gate_weight × activation_norm
• Top-26 frequency experts protected per layer
• See reap-prune-plan-full.json for per-expert scores