Running a trained model to produce output: the 'using it' phase, as opposed to training.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
Inference is what happens every time someone uses a model: the trained network runs forward on an input and produces output, one token at a time. No learning occurs; the weights are frozen.
The distinction from training matters economically. Training is a lumpy, capex-like cost paid up front; inference is the recurring cost of serving every request, forever. Falling inference cost per token is the load-bearing fact in most optimistic unit-economics arguments, and how far it falls, and who captures the savings, is a live dispute.