Definition
Model weights are learned numerical values that scale and combine signals inside a model. A neural-network layer may store them in a matrix or higher-dimensional tensor. During inference, operations use those values to transform an input into activations and, eventually, an output.
In strict usage, a weight multiplies one input or connection, while biases and some other learned values are separate parameters. In ordinary model discussions, "weights" often refers to the entire learned parameter set. A weight file or checkpoint serializes those tensors, sometimes together with optimizer state and metadata.
Training distributes behavior across many weights. One value does not usually correspond to one fact, rule, or source document. A model may memorize an example, but the checkpoint does not provide a readable table linking each output to the examples that caused it. Weight inspection and training-data provenance answer different questions.
Weights only have meaning with the architecture and configuration that use them. Tokenizer files, tensor names and shapes, numerical precision, and model code must match. Fine-tuning changes some or all weights; LoRA can store a separate low-rank update instead of another full copy.
Distinguish it from nearby terms
Parameters include weights, biases, and other learned values. Hyperparameters configure training or architecture but are not ordinarily learned by the same update rule. An activation is temporary state for one input. Source code and architecture define how all of these values interact.
Check your understanding
A checkpoint loads without errors after its tokenizer is replaced by one with a different vocabulary order. Why can valid tensor shapes still produce broken behavior, and which artifacts must be versioned together?