Uploaded August 2026 | Updated November 2026, 5 days ago
I tested the new MiniMax H3 W4A8 model in ComfyUI to see how it compares with INT8 in quality and speed. The W4A8 model is only around 12.5 GB, and in my test I could not see a clear quality difference from INT8.
I also tested Sol Attention, SageAttention, EasyCache, Turbo LoRA, the compressed VAE, and Ref2VA.
Some optimizations made generation much faster, but they also affected motion, quality, or audio. I show which settings worked well for me and which ones I would avoid in this workflow.
You need ComfyUI 0.31.0 or newer to use the W4A8 model.
Resources
MiniMax H3 W4A8:
huggingface.co/Kijai/MiniMax-H3-experimental
Sol Attention:
github.com/Saganaki22/ComfyUI-sol-attn
MiniMax H3 Turbo LoRA:
github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo
ComfyUI MiniMax H3 workflow:
aistudynow.com/minimax-h3-w4a8-in-comfyui-12-5-gb-model-tested
If you test W4A8, share your GPU and generation time in the comments.
#MiniMaxH3 #ComfyUI #AIVideo
I tested the new MiniMax H3 W4A8 model in ComfyUI to see how it compares with INT8 in quality and speed. The W4A8 model is only around 12.5 GB, and in my test I could not see a clear quality difference from INT8.
I also tested Sol Attention, SageAttention, EasyCache, Turbo LoRA, the compressed VAE, and Ref2VA.
Some optimizations made generation much faster, but they also affected motion, quality, or audio. I show which settings worked well for me and which ones I would avoid in this workflow.
You need ComfyUI 0.31.0 or newer to use the W4A8 model.
Resources
MiniMax H3 W4A8:
huggingface.co/Kijai/MiniMax-H3-experimental
Sol Attention:
github.com/Saganaki22/ComfyUI-sol-attn
MiniMax H3 Turbo LoRA:
github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo
ComfyUI MiniMax H3 workflow:
aistudynow.com/minimax-h3-w4a8-in-comfyui-12-5-gb-model-tested
If you test W4A8, share your GPU and generation time in the comments.
#MiniMaxH3 #ComfyUI #AIVideo


