Training adjusts a model's weights so that its computations better fit a learning objective. The resulting values are stored with the model and used during inference. They encode learned patterns rather than a simple readable collection of the original training examples.
Weights are only part of a working model system. The architecture, tokenization, runtime and input also affect behaviour. Parameter counts and storage precision influence memory requirements, but a larger file alone does not establish higher quality for every task.
ELI5
Model weights are like many adjustable settings that a program learns through practice. Together they influence what answer it produces.
For example, training may adjust those settings so a language model becomes better at predicting the next word. Using the trained model applies the learned settings to a new piece of text.
Are weights the same as a model's answers?
No. Weights are learned numerical parameters; answers are outputs computed from those parameters and the current input.
Why do weights use memory?
The runtime must store and access numerical values during computation, so their count and representation affect memory requirements.
