Quantization Method Consistency Inquiry
Are the quantization methods used for the model weights in your repository the same as the ones used in Kijai’s W4A8‑mixed weights below?
https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/minimax_h3_fl2va_pruned_w4a8_mixed.safetensors
Also, what are the differences compared to the quantization method in the repository linked below?
https://huggingface.co/Winnougan/ltx-2.5-w4a8-convrot-int4-convrot-Winnougan-Blessing/blob/main/diffusion_models/ltx-2.5-22b-dev-transformer-w4a8_convrot.safetensors
I’m trying to find a version that matches Kijai’s quantization, since I’ve tested his weights and they work really well.
Yeah, it's the same method as Kijai's, I tested it on minimax_h3 and it produced identical quantization metadata. From there, I just adapted the layer exclusion list for different architectures (currently LTX 2.5, Krea 2, and working on Ideogram 4). The main difference compared to Winnougan's is that theirs is a uniform/flat W4A8 across all blocks (~12.5GB), whereas mine calibrates against the BF16 model to keep sensitive transformer layers in higher precision to prevent output distortion. That's why this one is a bit larger (15.4GB), but the generation quality holds up better. I have not tested Winnougan's quants as of yet.
Yeah, it's the same method as Kijai's, I tested it on minimax_h3 and it produced identical quantization metadata. From there, I just adapted the layer exclusion list for different architectures (currently LTX 2.5, Krea 2, and working on Ideogram 4). The main difference compared to Winnougan's is that theirs is a uniform/flat W4A8 across all blocks (~12.5GB), whereas mine calibrates against the BF16 model to keep sensitive transformer layers in higher precision to prevent output distortion. That's why this one is a bit larger (15.4GB), but the generation quality holds up better. I have not tested Winnougan's quants as of yet.
Thank you so much — this is exactly the model weight I’ve been searching for. NVFP4 only accelerates inference on the 50‑series GPUs, but on my 4070 Ti, inference with NVFP4 weights is extremely slow. Meanwhile, the standard INT4 weights fall far short of W4A8‑mixed in terms of accuracy and quality.
Could you also create an unpruned W4A8‑mixed weight for the Minimax H3 model? Kijai quantized the pruned BF16 version into W4A8‑mixed, but in fact many people have been waiting for the unpruned W4A8‑mixed model weight.
https://huggingface.co/Kijai/MiniMax-H3-experimental/discussions/27
https://huggingface.co/Kijai/MiniMax-H3-experimental/discussions/17
https://huggingface.co/Kijai/MiniMax-H3-experimental/discussions/16
The unpruned is mainly for training, but I quantized it for you. I had to quantize the adaln_proj (what gets pruned) layers to int8 no convrot as it wasn't divisable by 256, degrading the quality. Leaving the layers in bf16 is essentially int8 model size, let me know how you find the quality https://huggingface.co/tsolful/Minimax_H3_W4A8/tree/main