Weights are the learned numerical parameters inside a machine learning model. In an LLM, they encode the patterns the model has absorbed during training and largely determine how it predicts the next token. When people say a model is “big,” they often mean it contains a huge number of weights that must be stored and loaded before inference can run.
In cyber security, weights matter because they are both valuable and sensitive. Attackers may try to steal them to copy a proprietary model, tamper with them to backdoor the system, or corrupt files so the model behaves unpredictably. Defenders protect weights with access controls, integrity checks, signed model artifacts, and careful handling of where they reside on disk, in RAM, or in VRAM. In local AI deployments, moving weights between memory tiers can make a model fit on limited hardware, but it also increases operational complexity and the need to secure those files.