The three learned projections of each token that attention uses to decide what to read from where.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
Attention gives every position three views of itself, each a learned projection of the same vector: a query for what this position is looking for, a key for what it offers to others, and a value for what it passes on when selected. Scores come from comparing queries against keys; the values are what actually get mixed.
The separation matters. A token can advertise something quite different from what it contributes, which is what lets attention route information rather than merely blend it.