aimosprite/nemotron-super-120b-a12b-nvfp4-count-reap-keep90

Nemotron Super NVFP4 Count-REAP Keep90

This repository contains a structurally pruned variant of nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4.

Pruning details

• Metric: count
• Keep ratio: 0.90
• Routed experts kept per MoE layer: 461 / 512
• Experts removed per MoE layer: 51 / 512
• Usage source: merged representative imo-answerbench final count dumps (pid61 + pid63)
• Materialization script: compress/reap/scripts/materialize_pruned_nemotron_model.py

Notes

• This is count-based REAP, not score-based REAP.
• reap_prune_plan.json records the exact per-layer expert remapping.