Before it
4-bit quantization decided which weights to protect using the weights themselves.
The problem
Weight magnitude is the wrong signal; what matters is which weights see large activations.
What it introduced
Use activation statistics to choose the salient channels to preserve, with no backpropagation needed.
Why it matters
Better accuracy at 4-bit and faster kernels, and one of the two quantization formats local tooling offers.
What modern models inherit
Activation-aware quantization in the serving stack.
What came next
Standard in vLLM and local inference alongside GPTQ and GGUF.