At a glance
Which variant should I pick?
Table with columns: Variant, Format, Link| Variant | Format | Link |
|---|
Nemotron-3-Super-64B | BF16 | link |
Nemotron-3-Super-64B-W4A16 | W4A16 | link |
Nemotron-3-Super-92B | BF16 | link |
Draft AutoRound quantization of a Nemotron Super checkpoint.
Base model: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
Draft status
This is a draft research release. It is published for inspection, reproducibility, and early runtime validation. It should not be treated as a final benchmarked production checkpoint.
How this was produced
We quantized the checkpoint with Intel AutoRound using the W4A16 scheme on the remote 8x RTX 3090 host. This lane is optimized for overnight completion and resumability rather than final accuracy tuning.
Settings used
- source checkpoint:
/mnt/llm_models/nemotron-super-compressions/nemotron_super_merged_long50_short15120_v2/reap_25pct
- source type:
REAP 25% pruned checkpoint
- quantizer:
intel/auto-round 0.10.2
- scheme:
W4A16
- format:
auto_round
- calibration dataset:
NeelNanda/pile-10k
- device_map:
auto
- nsamples:
128
- iters:
50
Notes
- upstream provenance is preserved through the base model link above
- this repo is intentionally marked draft while quantization/runtime validation is still in progress
- donation link added per maintainer request
License & citation
License inherited from the base model.
@misc{lasby2025reap,
title = {REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
author = {Mike Lasby and Ivan Lazarevich and Nish Sinnadurai and Sean Lie and Yani Ioannou and Vithursan Thangarasa},
year = {2025}, eprint = {2510.13999}, archivePrefix = {arXiv}
}
Made possible by NVIDIA · TNG Technology · Lambda · Prime Intellect · Hot Aisle.