What is inference?
The word for what happens every time you ask a model anything. Training built the mind; inference is using it.
Inference is running a trained model on your input to produce an answer. The model is a finished file of numbers; inference feeds your text through it and reads out the result, one step per generated piece. It is the difference between training, which happened once in a datacenter, and using, which happens every time you ask. Inference is light enough to run on a laptop, entirely offline.
Training built it; inference uses it.
Training is where the enormous cost lives: weeks of computation over oceans of text, producing the file of numbers that is the model. Inference is everything after. Your question goes in, the model's numbers run over it, and an answer comes out one piece at a time, each piece computed from the question and everything generated so far. The final pick among candidate words has a dial of its own; what temperature does covers that setting. That one-way split explains the shape of the whole industry: you will never train a model at home, and you do not need to. The same file, compressed, runs on hardware most people already own, because using a finished model is a far cheaper act than building one.
Asking changes nothing.
Inference reads the model and writes an answer. It does not rewrite the model. Whatever you tell it, whatever you paste in, the numbers on disk are the same before and after, which is worth knowing for two reasons. It is why a conversation does not make the model better, and why it does not make it yours: a local model cannot quietly learn from your documents any more than a cloud one can. What changes between sessions is only the context you supply, like the passages a retrieval loop places in front of it. The model itself stays exactly as it was built.
Why it runs on your machine.
Inference is arithmetic, and arithmetic does not need the internet. The file sits on your disk, the computation happens on your silicon, and no part of generating an answer is a network call. This is the technical fact under all the privacy promises: a tool that runs inference locally is not holding back your data under a policy you have to trust, it is structurally unable to send it anywhere. The connection matters only for what surrounds the model, such as fetching pages when you choose to or downloading the model itself. The answering step, the one that sees your words, works with the cable pulled out.