In plain words
Weights are parameters used to scale and combine values in a model’s calculations. Training adjusts them so the model’s outputs better match its objective.
A closer look
In a simple model, y = wx + b, the weight w determines how strongly input x affects the prediction, while b is a bias. Neural networks use large arrays of weights across many operations. Individual weights usually cannot be read as isolated facts or instructions; behavior emerges from their interactions with the architecture and input.
After training, weights are saved in a checkpoint and loaded to run inference. “Model weights” is also used loosely for the full collection of saved parameters. An open-weight release makes those values available under a license, but does not necessarily include the training data or all the code needed to reproduce training.
In practice
A simple price model may learn that floor area should have a positive influence on its estimate. In a language model, weights participate in far more complex transformations of token representations.
A useful distinction
Weights are not the same as the temporary attention scores computed for a particular input. And access to weights is not automatically unrestricted permission to use them: the release’s license still determines the terms.