---
title: MiniMax-H3-experimental
canonical_url: "https://www.modelscope.cn/models/Kijai/MiniMax-H3-experimental"
md_url: "https://www.modelscope.cn/models/Kijai/MiniMax-H3-experimental.md"
repository: Kijai/MiniMax-H3-experimental
last_updated: 2026-10-15
parameters: 84.8B
tensor_type:
  - F8_E4M3
  - BF16
  - F32
  - I64
  - I8
  - U8
  - F16
library_name:
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 12226
stars: 10
---

# MiniMax-H3-experimental

> MiniMax-H3-experimental - Kijai 在 ModelScope 开源的模型。w4a8 is a format using 4-bit weights and int8-convrot activations

- **Repository**: Kijai/MiniMax-H3-experimental
- **Parameters**: 84.8B
- **Downloads**: 12226
- **Stars**: 10
- **Last updated**: 2026-10-15

Source: https://www.modelscope.cn/models/Kijai/MiniMax-H3-experimental

---

w4a8 is a format using 4-bit weights and int8-convrot activations

Requires ComfyUI 0.31.0

---

int8_convrot VAE needs ComfyUI 0.36.0 for maximum speed benefit

It speeds up VAE decode times by ~2x

---

ref lora is the difference between fl2va and ref2va, completely experimental, I don't even know if it has a use case at this point.

---

w6a8 is a format using 6-bit weights and int8-convrot activation, work in progress

## minimax_h3_ref2va_pruned_w6a8_g32.safetensor

https://github.com/Comfy-Org/comfy-kitchen/pull/191

---

## minimax_h3_lynnreal_light_vae_int8_convrot.safetensors

https://huggingface.co/stdstu123/LynnReal-Onmi-light-vae

int8-convrot quant of the pruned light VAE, this is about 1.3x faster for decoding.

---

## PDMD loras

Converted and rank reduced from:

https://huggingface.co/pdmd2026/pdmd_2NFE_lora

https://huggingface.co/pdmd2026/pdmd_4NFE_lora

## DMAD lora

https://huggingface.co/ZhengmingYu/DMAD