MiniMax-H3-Image β W4A4
An experimental W4A4 quantization of MiniMax-H3-Image, created using PotatoForge.
This quantization was produced from the BF16 checkpoint published by Comfy-Org.
Why W4A4?
On my GTX 1660 Super (6 GB), W4A8 inference was nearly 4Γ slower, which motivated this W4A4 release.
The goal is to improve inference performance on older, resource-constrained GPUs while maintaining reasonable image quality (reasonable for an int4 quant).
License
This is an unofficial, modified version of MiniMax-H3-Image, originally developed by MiniMaxAI.
The model is subject to the MiniMax-H3-Image Community License Agreement, including its usage and redistribution restrictions.
Model tree for PotatoForge/Minimax-H3-Image-W4A4
Base model
MiniMaxAI/MiniMax-H3 Finetuned
MiniMaxAI/MiniMax-H3-Image Finetuned
Comfy-Org/MiniMax-H3-Image