# MiniMax H3 Singularity Pruned W4A8

## Summary

Pruned MiniMax H3 workflow packaged with a graft-preserving W4A8 conversion.

## Model Family

- `minimax`

## Status

- `imported-only`

## Source

- Model card: <https://huggingface.co/cicalooo/Minimax-h3_Singularity_v1.3_pruned_w4a8_graft_preserving>
- Workflow file: <https://huggingface.co/cicalooo/Minimax-h3_Singularity_v1.3_pruned_w4a8_graft_preserving/resolve/main/Minimax-h3_Singularity_v1.3_pruned_w4a8_graft_preserving.quality.json?download=true>
- Conversion notes: <https://huggingface.co/cicalooo/Minimax-h3_Singularity_v1.3_pruned_w4a8_graft_preserving/raw/main/Minimax-h3_Singularity_v1.3_pruned_w4a8_graft_preserving.quality.md>
- Original publisher: `cicalooo`
- Date imported: `2026-09-07`

## Developer notes

- The conversion notes describe a graft-preserving W4A8 pass over `Minimax-h3_Singularity_ref2va_v1.3_pruned_bf16.safetensors`.
- The workflow uses the MiniMax H3 long-video stack plus a low-step acceleration path.
- This is the same-day efficiency play, not a new art style drop.

## Our notes

- This is the MiniMax item most worth testing on a local card if you want a leaner path than full bf16.
- It is still not cheap on an RTX 3070 with `8 GB` VRAM, but it looks more practical than a plain full-size H3 pack.
- I would treat this as a serious candidate, not just a watch item.

## Required custom nodes

- `Smite79/MiniMax-H3-LongVideos`
- current ComfyUI MiniMax-H3 support
- `ComfyUI-MiniMax-H3-PDD-Acc` for the acceleration path used in the workflow

## Required models

- `Minimax-h3_Singularity_ref2va_v1.3_pruned_w4a8_graft_preserving.safetensors`
- MiniMax H3 video VAE
- MiniMax H3 audio VAE
- matching MiniMax H3 text encoder

## Notes

- Best used as the leaner local comparison point against the official or full-size H3 templates.