mixed precision
Quantizing different parts of a model to different bit depths instead of crushing all of it the same amount. A typical recipe holds the roughly 15% of load-bearing weights at 8 bits and pushes the other 85% down to 4, which costs a little more than uniform 4-bit and keeps most of what uniform 4-bit throws away.
Not all weights matter equally. Damage some and the model barely notices; damage others and it stops being able to hold a thought. Uniform compression has no way to tell them apart, so it pays the same price everywhere.
The photograph on the quantization page shows the same idea. Turn mixed precision on and the sun and the ridge stay at 8 bits while the sky drops to whatever the slider says. For nearly the same file size, the banding lands where nobody was looking.
The frontier version of this is asymmetric: DwarfStar 4 compresses only the routed experts of DeepSeek V4 Flash to 2 bits and leaves attention and the shared components high, which is what lets a 284B model run on a 128 GB Mac.