Precision is the number format a model's math uses, such as INT8 or 16-bit. Each number inside an AI model is stored with a certain number of bits, and fewer bits mean less detail but also less memory and faster calculation.

It matters because it affects how large a model is, how quickly it runs and how good its answers are. Running a model at lower precision is one of the main ways to fit it on a phone or laptop, and chip ratings such as TOPS are tied to a specific precision.

Lower precision is not free. It can reduce accuracy, and the effect varies with the model, the task and the method of conversion. Comparisons are meaningful only when the precision is the same, because a chip may be much faster at one format than another, and a figure quoted at a low precision may not apply at a higher one.

When reading a spec or a result, check the precision it refers to, test the quality of the model at that setting on your own tasks and compare like with like.