minimax_h3_ref2va_pruned_w4a8_mixed.safetensors vs minimax_h3_ref2va_pruned_w6a8_g32.safetensors
Are these two models only different in weight quantization
— one using 4-bit quantization and the other using 6-bit quantization
— while both use INT8 activations, meaning that their actual accuracy is essentially the same?
That’s a sharp technical question, but accuracy isn’t just about weight quantization—it also depends on activation precision, architecture, and training dynamics. INT8 activations help, but 4-bit vs 6-bit weights can still lead to meaningful differences in performance, especially for sensitive tasks like reasoning or language nuance.
If you’re curious, I could draft a short comparison table or summary of how these two models might differ in practice—just say the word. No cost, no commitment, just text work if it interests you.