Why is local AI slow on my laptop?
The wait you feel has three parts, and only one of them is your hardware's fault. None of them change what the answer says.
A local model runs slow for three main reasons: the model is large, the memory feeding it is slow, and generation is sequential, one token at a time. Your CPU can still run small models, just at a few tokens per second instead of dozens. The fixes that work are a smaller or more quantized model and, when answers run long, an entry-level GPU. Slow never changes the answer's quality.
Generation is the part that waits.
Search, indexing and retrieval are fast on any machine; the wait you feel is almost always generation. A model writes one token at a time, each computed from the question and everything written so far, so speed is measured in tokens per second and the whole answer takes as long as its slowest stretch. Reading your documents back is quick. Writing the answer out is the expensive half, and it repeats with every question you ask.
What the hardware is actually doing.
Each token is a pile of small matrix multiplications across every parameter the model has, so a bigger parameter count costs speed directly. The other limit is memory bandwidth: the arithmetic units spend most of their time waiting for the model's numbers to arrive, which is why a GPU helps more than its raw compute suggests and why a laptop CPU, doing a few multiplications at a time, lands at a few tokens per second instead of dozens. That is the whole difference between the two, and it is laid out on what a GPU actually buys.
What actually helps.
Three changes work. Drop the model size, since a 7B class model is the sweet spot for a laptop and something twice as large is not. Push quantization down a step, because a 4-bit file runs quicker than an 8-bit one at little quality cost, covered on what quantization does. And add even an entry-level GPU if long answers are your normal work. What does not help: closing browser tabs. The model is the consumer, not your background apps, and the budget that actually matters is memory, covered on how much RAM a local AI needs.