Quantization
Shrinking a model's internal numbers down to lower precision so it runs faster and fits in far less memory — often the difference between needing a data center and running on a phone.
A model trained with high-precision numbers (32-bit) can be converted to use lower-precision ones (16-, 8-, even 4-bit) after training. Each number takes less memory and less time to compute, so the same model runs faster and fits on cheaper hardware — usually at a small, often barely noticeable, cost in accuracy.
Quantization is the main reason a capable AI model can now run entirely on a phone or laptop with no server round-trip at all, and it's one of several efficiency techniques — alongside distillation and mixture-of-experts — that has narrowed the gap between what a small local model and a giant cloud model can actually do, which matters directly for the memory-supply crunch that HBM demand is built on.