# MiniMax H3 PlagueKind V2

## Summary

`MiniMax H3` workflow focused on ease of use, int8 convrot pieces, and a lower-memory path for local testing.

## Model Family

- `minimax`

## Status

- `imported-only`

## Source

- Model page: <https://huggingface.co/Plaguekind/MiniMax-H3>
- Workflow file: <https://huggingface.co/Plaguekind/MiniMax-H3/resolve/main/PlagueKind-MinimaxH3-V2.json>
- Original publisher: `Plaguekind`
- Date imported: `2026-08-09`

## Developer notes

- The model card says V2 adds an int8 convrot video VAE, memory chunks, and a Kijai resized turbo LoRA.
- The author notes the workflow was already near the cap for `16 GB` VRAM, so this is not a tiny setup.
- It requires several community custom node packs rather than only stock ComfyUI.

## Our notes

- This is one of the more serious MiniMax releases of the day because it is clearly trying to make H3 easier to run, not just more flashy.
- On an `RTX 3070 / 8 GB`, it still looks heavy, but the memory-saving tricks make it worth keeping on the radar.
- If MiniMax starts feeling close to practical, this is a stronger watch item than the generic wrappers.

## Required custom nodes

- `ComfyUI-PlagueKind-Nodes`
- `ComfyUI-KJNodes`
- `rgthree-comfy`
- `Nvidia_RTX_Nodes_ComfyUI`
- `ComfyUI-VideoHelperSuite`
- `SageAttention`

## Required models

- `Comfy-Org/MiniMax-H3`
- `Kijai/MiniMax-H3-experimental`
- `Comfy-Org/frame_interpolation`
- `Kijai/MiniMax-H3_comfy`

## Notes

- Best same-day MiniMax workflow for a more tuned local-test path.